Deep learning-based traditional Chinese medicine tongue diagnosis image intelligent analysis system

Abstract

The invention relates to the technical field of artificial intelligence and traditional Chinese medicine diagnosis, in particular to a traditional Chinese medicine tongue diagnosis image intelligent analysis system based on deep learning, which comprises a tongue picture acquisition module, a tongue picture measurement module, a tongue picture comparison module, a tongue picture comprehensive analysis module, a health consultation module and a data management module. Accurate extraction and quantitative analysis of tongue picture color features are realized by constructing a color manifold space and introducing Riemannian metric tensor and geodesic distance calculation; performing dynamic evolution tracking on tongue image change by adopting a thermal equation and a curvature flow model on a manifold; based on the geodesic distance, constructing a mapping relation between color features and syndromes, and generating syndrome probability distribution; meanwhile, tongue images collected at different times are compared and analyzed, the treatment effect is evaluated, the system is further combined with a traditional Chinese medicine knowledge graph, syndrome diagnosis, disease interpretation and dialectical treatment suggestions are provided, and the limitation that traditional tongue diagnosis is high in subjectivity and lacks quantitative standards is broken through.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and traditional Chinese medicine diagnostic technology, specifically to a deep learning-based intelligent analysis system for tongue diagnosis images in traditional Chinese medicine, used to achieve intelligent acquisition, analysis, diagnosis, and health assessment of tongue images. Background Technology

[0002] Tongue diagnosis in Traditional Chinese Medicine (TCM), as an important component of inspection (one of the four diagnostic methods in TCM, along with auscultation, inquiry, and palpation), is based on the theory that the tongue reflects the heart and mirrors the internal organs. By observing the tongue's appearance (including its texture, coating, and shape), one can reflect the function of the internal organs, the state of Qi, blood, and body fluids, as well as the nature and severity of diseases. Traditional tongue diagnosis relies heavily on the physician's subjective experience, which presents several problems: first, the diagnostic results lack objectivity and repeatability, with different physicians potentially interpreting the same tongue appearance differently; second, it lacks quantitative standards, making it impossible to accurately measure subtle changes in tongue characteristics; and third, it cannot achieve automatic comparison and analysis of historical tongue data, hindering the objective evaluation of treatment effectiveness.

[0003] Existing technologies include several image-processing-based tongue diagnosis assistance systems, but these systems typically focus only on simple color analysis or morphological recognition, lacking in-depth exploration of the relationship between tongue features and TCM syndromes. Furthermore, traditional image processing methods have limitations in handling complex and variable tongue features, particularly in recognizing subtle color changes and complex textures. In addition, existing systems generally lack the ability to dynamically track changes in tongue appearance, failing to provide objective evidence for evaluating the evolution of TCM syndromes and the effectiveness of treatment.

[0004] With the development of deep learning technology, especially the remarkable achievements of convolutional neural networks (CNNs) in the field of medical image processing, new technical means have been provided for intelligent tongue diagnosis. However, how to organically combine advanced mathematical theories with traditional Chinese medicine theories to construct an intelligent tongue diagnosis analysis system that conforms to both Chinese medicine theory and modern scientific foundation remains an urgent problem to be solved. Summary of the Invention

[0005] The purpose of this invention is to provide a deep learning-based intelligent image analysis system for TCM tongue diagnosis. By combining differential geometry theory with deep learning technology, it can accurately extract and analyze tongue features, establish a scientific mapping relationship between tongue features and TCM syndromes, and improve the objectivity, accuracy and repeatability of tongue diagnosis.

[0006] This invention proposes a deep learning-based intelligent image analysis system for traditional Chinese medicine tongue diagnosis, comprising:

[0007] The tongue image acquisition module is used to acquire standardized tongue images, evaluate the amount of tongue image information, image quality, tongue image location and clarity, and generate a tongue image acquisition quality score.

[0008] The tongue image measurement module is connected to the tongue image acquisition module and is used to receive the tongue image acquired by the tongue image acquisition module, construct a color manifold space based on the tongue image, calculate the geodesic distance in the color manifold space, establish the mapping relationship between color features and syndromes, perform dynamic evolution analysis of color features, and generate quantitative analysis results of tongue color and tongue coating distribution.

[0009] The tongue image comparison module is connected to the tongue image measurement module and is used to receive the quantitative analysis results generated by the tongue image measurement module, compare and analyze tongue images collected at different times, and generate a report on changes before and after treatment.

[0010] The tongue image comprehensive analysis module is connected to the tongue image measurement module and the tongue image comparison module. It is used to receive the quantitative analysis results of the tongue image measurement module and the change report of the tongue image comparison module, to conduct a multi-dimensional comprehensive assessment of health status, and to generate quantitative scores and treatment suggestions.

[0011] The health consultation module is connected to the tongue image comprehensive analysis module. It is used to receive the evaluation results of the tongue image comprehensive analysis module, query the traditional Chinese medicine knowledge graph, compare the diagnosis results with related symptoms, and generate traditional Chinese medicine diagnosis results, disease explanations, and etiological differentiation.

[0012] Preferably, the tongue image measurement module includes:

[0013] Color manifold construction unit is used to reconstruct the RGB color space in the tongue image into a color manifold with Riemannian metric, construct the tongue color distribution density function, and calculate the manifold curvature features;

[0014] The geodesic distance calculation unit is connected to the color manifold construction unit and is used to solve the geodesic equation on the color manifold, construct the geodesic distance matrix, and calculate the intrinsic distance between color feature points and symptom reference points.

[0015] The syndrome mapping analysis unit is connected to the geodesic distance calculation unit and is used to calculate the syndrome similarity score and generate the syndrome probability distribution based on the geodesic distance matrix.

[0016] The dynamic evolution tracking unit, connected to the color manifold construction unit and the syndrome mapping analysis unit, is used to analyze the changing patterns of the color manifold over time, predict the evolution trend of syndromes, and evaluate the treatment effect.

[0017] Preferably, the color manifold building unit includes:

[0018] The coordinate system reconstruction module is used to define the local coordinate system of the color manifold, establish mapping functions, and ensure smooth transitions between regions.

[0019] The metric tensor construction module, connected to the coordinate system reconstruction module, is used to define the Riemann metric tensor on the manifold, optimize the sensitivity of the metric to the syndrome features, and ensure the continuity of the metric.

[0020] The manifold feature extraction module, connected to the metric tensor construction module, is used to calculate the curvature features of the manifold, identify topological features, and integrate regional features.

[0021] Preferably, the geodesic distance calculation unit includes:

[0022] The mesh structure design module is used to build a mesh structure that adapts to tongue images, mapping pixels to mesh nodes on a manifold;

[0023] The geodesic equation solving module, connected to the grid structure design module, is used to discretize the continuous geodesic equation, realize adaptive time step control, and solve the geodesic equation.

[0024] The key point selection module is connected to the geodesic equation solving module and is used to extract representative color points, identify sensitive areas of syndrome, and control sampling density.

[0025] The distance matrix construction module, connected to the key point selection module and the geodesic equation solving module, is used to calculate the geodesic distance between key points, interpolate the distance values ​​of non-key points, and optimize the storage and retrieval of the distance matrix.

[0026] Preferably, the syndrome mapping analysis unit includes:

[0027] The syndrome feature space construction module is used to define the syndrome feature representation framework, integrate expert knowledge, and optimize syndrome templates;

[0028] The basic mapping design module, connected to the syndrome feature space construction module, is used to design the mapping function from color manifold features to the syndrome space and learn complex nonlinear mapping relationships.

[0029] A multi-scale similarity calculation module, connected to the basic mapping design module, is used to calculate the similarity of local regions, aggregate global similarity, and perform context-sensitive adjustments.

[0030] The syndrome probability distribution generation module is connected to the multi-scale similarity calculation module and is used to convert similarity into probability values, analyze syndrome combination patterns, and quantify diagnostic uncertainty.

[0031] Preferably, the dynamic evolution tracking unit includes:

[0032] The time-series data preprocessing module is used to align tongue image data from multiple time points, normalize timestamps, and handle abnormal data.

[0033] The thermal equation simulation module, connected to the time-series data preprocessing module, is used to solve the thermal equation on the manifold and analyze the diffusion process of color features.

[0034] The curvature flow analysis module, connected to the time-series data preprocessing module, is used to simulate the evolution of manifold curvature flow, track the trajectory of feature points, and predict the steady state.

[0035] The treatment effect evaluation module, connected to the thermal equation simulation module and the curvature flow analysis module, is used to identify treatment response patterns, build an evolution path library, compare actual trajectories with reference paths, and generate treatment suggestions.

[0036] Preferably, the tongue image comparison module includes:

[0037] The tongue color comparison and analysis unit is used to receive the quantitative analysis results of the tongue image measurement module and compare and analyze the color features of the tongue image photos collected before and after treatment.

[0038] The tongue coating quality comparison and analysis unit is used to receive the quantitative analysis results of the tongue image measurement module, compare the tongue coating quality information collected before and after treatment, and output the symptom changes.

[0039] The tongue morphology comparison and analysis unit is used to receive the quantitative analysis results of the tongue image measurement module, analyze the changes in tongue morphology before and after treatment, and evaluate the treatment effect.

[0040] Preferably, the tongue image comprehensive analysis module includes:

[0041] The multi-dimensional feature fusion unit is used to receive the quantitative analysis results of the tongue image measurement module and the change report of the tongue image comparison module, and integrate the analysis results of tongue color, tongue coating and tongue morphology;

[0042] A health status assessment unit, connected to the multidimensional feature fusion unit, is used to calculate a quantitative health status score based on the integrated analysis results and according to a preset scoring standard.

[0043] The treatment plan generation unit is connected to the health status assessment unit and is used to generate personalized treatment suggestions based on the health status assessment results and traditional Chinese medicine theory.

[0044] Preferably, the health consultation module includes:

[0045] The knowledge graph query unit is used to receive the evaluation results of the tongue image comprehensive analysis module and retrieve relevant syndrome information in the traditional Chinese medicine knowledge graph;

[0046] The diagnosis result generation unit is connected to the knowledge graph query unit and is used to generate traditional Chinese medicine diagnosis results based on the knowledge graph retrieval results and tongue image analysis results.

[0047] The symptom interpretation unit, connected to the diagnosis result generation unit, is used to provide symptom interpretations in traditional Chinese medicine theory based on the diagnosis results.

[0048] The etiology dialectics unit, connected to the diagnosis result generation unit, is used to analyze the etiology based on the diagnosis result and patient information, and provide dialectical treatment suggestions.

[0049] Preferably, the system also includes a data management module, which is connected to the tongue image acquisition module, the tongue image measurement module, the tongue image comparison module, the tongue image comprehensive analysis module, and the health consultation module. The data management module is used to store historically acquired tongue image data, analysis results, and diagnostic records, manage patient information, support data retrieval and comparison, and provide the data foundation for system operation.

[0050] The beneficial effects of this invention include:

[0051] 1. Standardization and quality assessment of tongue image collection were achieved, ensuring the quality and reliability of the analytical data;

[0052] 2. By constructing a color manifold space and calculating geodesic distance, the limitations of the traditional RGB color space are overcome, and subtle differences in tongue color features are captured more accurately, thus improving the accuracy of syndrome recognition.

[0053] 3. Based on the dynamic evolution analysis of differential geometry, the changes in tongue appearance were accurately tracked, providing an objective quantitative basis for the evaluation of treatment effects;

[0054] 4. Through multi-dimensional feature fusion and comprehensive analysis, combined with TCM knowledge graph, intelligent reasoning from symptom presentation to syndrome diagnosis is realized, improving the comprehensiveness and accuracy of diagnosis;

[0055] 5. A complete tongue image data management system has been established, supporting historical data comparison and analysis, and providing data support for the formulation of personalized treatment plans.

[0056] This invention organically combines modern mathematical theory, deep learning technology and traditional Chinese medicine theory, which not only improves the scientificity and reliability of tongue diagnosis in Chinese medicine, but also provides a new technical path for the inheritance, development and innovation of Chinese medicine theory. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the overall system architecture of the present invention;

[0058] Figure 2 This is a schematic diagram of the tongue image measurement module of the present invention;

[0059] Figure 3 This is a flowchart of the color manifold construction unit of the present invention;

[0060] Figure 4 This is a flowchart of the geodesic distance calculation unit of the present invention;

[0061] Figure 5 This is a flowchart of the syndrome mapping analysis unit of the present invention;

[0062] Figure 6 This is a flowchart of the dynamic evolution tracking unit of the present invention;

[0063] Figure 7 This is a flowchart of the tongue image comparison module of the present invention;

[0064] Figure 8 This is a flowchart of the tongue image comprehensive analysis module of the present invention;

[0065] Figure 9 This is a flowchart of the health consultation module of the present invention;

[0066] Figure 10 This is a structural diagram of the data management module of the present invention. Detailed Implementation

[0067] Please refer to Figures 1-10 The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0068] Reference Figure 1 The present invention provides a deep learning-based intelligent analysis system for TCM tongue diagnosis images, comprising a tongue image acquisition module 1, a tongue image measurement module 2, a tongue image comparison module 3, a tongue image comprehensive analysis module 4, a health consultation module 5, and a data management module 6.

[0069] The tongue image acquisition module 1 is used to acquire standardized tongue images, evaluate the amount of tongue image information, image quality, tongue position, and clarity, and generate a tongue image acquisition quality score. In practical applications, the tongue image acquisition module 1 uses standardized light source equipment (preferably an LED cold light source with a color temperature of 5500K) and an image acquisition device at a fixed distance (preferably 30cm) to ensure a consistent acquisition environment. The system provides a standard tongue image reference sample, which the user can use to adjust the acquisition posture and angle. The tongue image acquisition module 1 analyzes the quality of the acquired images in real time, evaluating parameters such as image clarity, brightness, contrast, and tongue position, and generating a quality score of 0-100. When the score is lower than a set threshold (preferably 75 points), the system provides improvement suggestions, such as adjusting the lighting or placing the tongue in the center of the image.

[0070] The tongue image measurement module 2 is connected to the tongue image acquisition module 1 and is used to receive the tongue image acquired by the tongue image acquisition module 1. Based on the tongue image, a color manifold space is constructed, geodesic distances are calculated in the color manifold space, a mapping relationship between color features and syndromes is established, dynamic evolution analysis of color features is performed, and quantitative analysis results of tongue color and tongue coating distribution are generated.

[0071] Reference Figure 2 The tongue image measurement module 2 includes a color manifold construction unit 21, a geodesic distance calculation unit 22, a syndrome mapping analysis unit 23, and a dynamic evolution tracking unit 24.

[0072] Reference Figure 3 The color manifold construction unit 21 is used to reconstruct the RGB color space in the tongue image into a color manifold with Riemannian metric, construct the tongue color distribution density function, and calculate the manifold curvature features. In a preferred embodiment of the present invention, the color manifold construction unit 21 first segments the tongue region and extracts the RGB values ​​of the effective tongue pixels. Then, the RGB color space is mapped onto a three-dimensional Riemannian manifold M, which has an intrinsic metric structure and can more accurately describe the actual visual differences between colors.

[0073] Specifically, the color manifold construction unit 21 includes a coordinate system reconstruction module 211, a metric tensor construction module 212, and a manifold feature extraction module 213.

[0074] The coordinate system reconstruction module 211 defines the local coordinate system of the color manifold M and establishes a mapping function to ensure smooth transitions between regions. In an embodiment of the present invention, the local coordinate system on the manifold M can be represented as follows: ,in To cover the open set, Let be a mapping function that maps points in RGB space onto the manifold M. Where: Denotes the first on manifold M An open set forms an open cover of the manifold; Indicates starting from the opening set A homeomorphism to Euclidean space is used to achieve local coordinate representation. To enhance the system's sensitivity to different syndrome characteristics, the coordinate transformation matrix is ​​optimized so that color changes related to specific syndromes are more significant in the transformed coordinates.

[0075] The metric tensor construction module 212 defines the Riemann metric tensor g on the manifold M, optimizing the sensitivity of the metric to symptom features and ensuring the continuity of the metric. The Riemann metric tensor g defines the distance metric between any two points on the manifold, in the form:

[0076] ,

[0077] in: To measure the i-th row and j-th column component of a tensor, is a A symmetric positive definite matrix; This represents the k-th component of the color space (corresponding to the RGB value). The red "R" indicates red. The green "G" indicates... (Represents blue B); and For local coordinates on the manifold, ; These are weighting coefficients used to adjust the importance of each color component, and they satisfy... and ; Represents color components coordinates The partial derivative reflects the rate of change of color in that direction. In practice, for different syndrome types, The values ​​will vary; for example, for heat syndromes, the weight of the red component... Typically set higher (e.g.) , For cold syndromes, the weight of blue or white components will increase accordingly.

[0078] The manifold feature extraction module 213 calculates the curvature features of the manifold, identifies topological features, and integrates region features. The curvature features of the manifold are an important indicator characterizing the intrinsic geometric properties of color distribution. The system calculates the Riemann curvature tensor. Its derived scalar curvature R serves as a characteristic indicator of tongue color syndrome:

[0079] ,

[0080] in: Scalar curvature is the numerical representation of the curvature at each point on the manifold; For measuring tensors The elements of the inverse matrix satisfy ,in For the Kronecker delta notation, when The value is 1 if it is true, and 0 otherwise. For the Ricci curvature tensor, by the Riemann curvature tensor The result of the contraction is that , representing the curvature characteristics of a manifold in different directions. In practical applications, the distribution characteristics of the scalar curvature R value are significantly correlated with specific syndromes. For example, the tongue appearance of damp-heat syndrome usually shows a high curvature value (R>0.75) in a specific region, while qi deficiency syndrome shows a relatively flat curvature distribution. .

[0081] Reference Figure 4The geodesic distance calculation unit 22 is used to solve the geodesic equation on the color manifold, construct the geodesic distance matrix, and calculate the intrinsic distance between color feature points and symptom reference points. A geodesic is the shortest path between two points on a manifold, and its length reflects the true visual difference between color features, being more accurate than Euclidean distance.

[0082] The geodesic distance calculation unit 22 includes a grid structure design module 221, a geodesic equation solving module 222, a key point selection module 223, and a distance matrix construction module 224.

[0083] The grid structure design module 221 constructs a grid structure adapted to the tongue image, mapping pixels to grid nodes on the manifold. In practice, the system adopts an irregular triangular grid structure, and the grid density is adaptively adjusted according to the intensity of color changes in the tongue image. The grid density is higher (approximately 100 nodes / square centimeter) in areas with obvious color changes (such as the tip and sides of the tongue), while the grid density is lower (approximately 30 nodes / square centimeter) in areas with relatively uniform color (such as the middle of the tongue).

[0084] The geodesic equation solving module 222 discretizes the continuous geodesic equations, implements adaptive time step control, and solves the geodesic equations. The mathematical form of the geodesic equation is:

[0085] ,

[0086] in: This represents the k-th coordinate component on the manifold. ; The path parameter describes the location of a point on a geodesic line, and its value typically ranges from [0,1]. Corresponding to the starting point, Corresponding endpoint; Representing coordinates For parameters The second derivative of , i.e., acceleration; and These represent the i-th and j-th coordinate components relative to the parameters, respectively. The first derivative of , i.e., velocity; The Christoffel notation represents a connection on a manifold, defined as follows: ,in To measure the inverse of a tensor, The components are metric tensors. In actual calculations, the system uses an improved FastMarching algorithm to solve the discretized geodesic equations, with an algorithm complexity controlled at O(NlogN), where N is the number of grid nodes.

[0087] The key point selection module 223 extracts representative color points, identifies syndrome-sensitive areas, and controls sampling density. To optimize computational resources, the system does not calculate the geodesic distance between all pixels in the tongue image, but intelligently selects a representative set of points. The system uses a color clustering algorithm to extract representative points, identifying approximately 100-200 key points on a standard tongue image. Simultaneously, based on traditional Chinese medicine theory, the sampling density is increased in syndrome-sensitive areas (such as the tip of the tongue corresponding to heart fire, and the sides of the tongue corresponding to liver and gallbladder).

[0088] The distance matrix construction module 224 calculates the geodesic distances between keypoints, interpolates the distance values ​​of non-keypoints, and optimizes the storage and retrieval of the distance matrix. The system constructs an n×n geodesic distance matrix D (where n is the number of keypoints), where... This represents the geodesic distance between key points i and j. For non-key points, the system uses a weighted interpolation method to estimate their distance values.

[0089] ,

[0090] in: This represents the geodesic distance between point p and point q; For the i-th key point near point p, usually the k nearest neighbors of point p are taken; k is the number of nearest neighbors to consider, usually a value of 5; These are the weighting coefficients, related to point p to... It is inversely proportional to the Euclidean distance, defined as ,in Indicate that point p and The Euclidean distance between points, with weighting coefficients satisfying .

[0091] Reference Figure 5 The syndrome mapping analysis unit 23 is used to calculate syndrome similarity scores and generate syndrome probability distributions based on the geodesic distance matrix. The syndrome mapping analysis unit 23 establishes a mapping relationship between color features and traditional Chinese medicine syndromes, realizing the transformation from objective features to diagnostic results.

[0092] The syndrome mapping analysis unit 23 includes a syndrome feature space construction module 231, a basic mapping design module 232, a multi-scale similarity calculation module 233, and a syndrome probability distribution generation module 234.

[0093] The syndrome feature space construction module 231 defines the syndrome feature representation framework, integrates expert knowledge, and optimizes syndrome templates. The system constructs a standard syndrome template library based on large-scale clinical data and expert knowledge. Each syndrome (such as Qi deficiency, Blood deficiency, Yin deficiency, Yang deficiency, Damp-heat, etc.) corresponds to a set of feature vectors, describing the color, shape, and texture characteristics of the typical tongue appearance for that syndrome. The template library is continuously optimized through clinical validation to ensure its representativeness and accuracy.

[0094] The basic mapping design module 232 designs a mapping function from color manifold features to the syndrome space, learning complex nonlinear mapping relationships. The system employs a deep neural network structure to learn the complex mapping relationship between color manifold features and syndromes. The network architecture consists of three main parts: feature extraction, feature fusion, and syndrome mapping. The feature extraction part uses a multi-layer convolutional structure to extract local and global color features; the feature fusion part integrates features from different scales and regions; and the syndrome mapping part maps the fused features to the syndrome space.

[0095] The multi-scale similarity calculation module 233 calculates the similarity of local regions, aggregates global similarity, and performs context-sensitive adjustments. The system implements multi-scale syndrome similarity calculation, focusing on both local regional features (such as the tip, middle, and root of the tongue) and the overall characteristics of the global tongue image. Local similarity is calculated based on geodesic distance.

[0096] ,

[0097] in: For the local region x and the symptom reference point The similarity score ranges from (0,1], with a larger value indicating a higher similarity. This is the geodesic distance between the two. This is a scaling factor (usually set to 0.1-0.5), which controls the sensitivity of similarity to changes in distance; The natural exponential function is used. Global similarity is obtained by weighted aggregation of local similarity:

[0098] ,

[0099] in: Tongue appearance X and symptoms The global similarity is defined, with values ​​ranging from (0,1]. Represents a complete image of the tongue; Indicates the i-th syndrome; This refers to the j-th local region of the tongue image X, such as the tip, middle, or root of the tongue. This represents the total number of local regions to be divided, typically ranging from 5 to 7. For symptoms Reference features in the j-th local region; Let be the weight of the j-th region, representing the importance of that region in the overall diagnosis, and satisfying the following conditions: In practice, the weight of different areas is adjusted according to the syndrome type. For example, for heart fire syndrome, the tip of the tongue has a higher weight (about 0.4-0.5); for spleen and stomach syndrome, the middle of the tongue has a higher weight (about 0.4-0.5).

[0100] The syndrome probability distribution generation module 234 converts similarity into probability values, analyzes syndrome combination patterns, and quantifies diagnostic uncertainty. The system converts similarity into a syndrome probability distribution, reflecting the likelihood of each syndrome.

[0101] ,

[0102] in: Indicates the syndrome under a given tongue image X The probability of is in the range [0,1] and satisfies . ; Represents a tongue image; Indicates the i-th syndrome; Tongue appearance X and symptoms Global similarity; The temperature parameter (usually set to 1) controls the "sharpness" of the probability distribution; a smaller value... The value makes the probability distribution more concentrated on highly similar symptoms; The total number of syndrome categories identified by the system; It is a natural exponential function; This represents the summation of all n syndromes. The system also analyzes syndrome combination patterns and identifies common syndrome coexistence relationships (such as Qi and Yin deficiency, liver stagnation and spleen deficiency, etc.), improving the comprehensiveness of diagnosis.

[0103] In addition, the system quantifies the uncertainty of the diagnosis and calculates the reliability index:

[0104] ,

[0105] in: The value ranges from [0,1] to indicate the reliability of the diagnostic result; a larger value indicates a more certain diagnostic result. Let P be the entropy of the probability distribution, calculated using the following formula: , indicating the uncertainty of the probability distribution; Indicating symptoms The probability of; is the natural logarithm of n, representing the maximum possible value of entropy, used for normalization. A higher confidence level (closer to 1) indicates a more certain diagnosis; a lower confidence level (closer to 0) indicates the need for more information or human intervention.

[0106] Reference Figure 6 The dynamic evolution tracking unit 24 is used to analyze the changes in color manifold over time, predict the evolution trend of syndromes, and evaluate the treatment effect. This unit is one of the core innovations of this system, breaking through the limitations of traditional tongue diagnosis that only focuses on static features, and realizing the quantitative analysis of dynamic changes in tongue appearance.

[0107] The dynamic evolution tracking unit 24 includes a time-series data preprocessing module 241, a thermal equation simulation module 242, a curvature flow analysis module 243, and a treatment effect evaluation module 244.

[0108] The time-series data preprocessing module 241 aligns tongue image data from multiple time points, normalizes timestamps, and handles outlier data. The system processes tongue image data collected from patients at different time points, ensuring consistency in tongue position through image registration technology. For data with irregular sampling intervals, the system performs time normalization processing, converting it into an evenly spaced time series. Simultaneously, the system detects and processes outliers, such as color shifts caused by differences in lighting conditions.

[0109] The thermal equation simulation module 242 solves the thermal equations on the manifold to analyze the diffusion process of color features. The system constructs a thermal equation model on the color manifold to describe the diffusion process of color features on the manifold:

[0110] ,

[0111] in: Let x represent the color feature distribution function, which is a scalar function defined on the manifold M and varies with position x and time t. This is a time parameter, representing the time progression of the diffusion process; It represents the partial derivative of u with respect to time t, describing the rate of change of u with time; Let be the Laplace-Beltrami operator on manifold M, a second-order differential operator in Riemannian geometry, defined as follows: ,in To measure the determinant of tensor g, Let g be the inverse matrix element of g. The system solves the partial differential equation numerically and analyzes the variation of color characteristics over time. Heat kernels at different time scales. It reveals different patterns of color change: small time scales (t=0.1-0.5) reflect local changes, while large time scales (t=1-5) reflect global trends.

[0112] The curvature flow analysis module 243 simulates the evolution of curvature flow in a manifold, tracks the trajectories of characteristic points, and predicts the steady state. The system constructs the Riemannian curvature flow equations to simulate the natural evolution of the manifold structure over time.

[0113] ,

[0114] in: To measure the tensor components, it is a 3×3 symmetric positive definite matrix that describes the distance metric on the manifold M; This represents the rate of change of a metric tensor over time. For time parameters; The Ricci curvature tensor, derived from the Riemann curvature tensor contraction, describes the curvature properties of the manifold in various directions. The system analyzes the trajectories of feature points under curvature flow, capturing the main directions and rates of color change. Eigenvalue evolution analysis is then used to further investigate this. The system can predict the stable state of color changes, providing a basis for predicting treatment effects. Among these, Represents the metric tensor The i-th eigenvalue at time t .

[0115] The treatment efficacy evaluation module 244 identifies treatment response patterns, constructs an evolutionary path library, compares actual trajectories with reference paths, and generates treatment suggestions. The system builds a standard evolutionary path library based on large-scale clinical data, recording the tongue appearance changes of different syndromes under typical treatment plans. For example, under heat-clearing and dampness-draining treatment, the tongue appearance of damp-heat syndrome typically follows an evolutionary path: red tongue with yellow, greasy coating → red tongue with thin yellow coating → pale red tongue with thin white coating. The system compares the patient's actual tongue appearance changes with the standard path and calculates the trajectory similarity.

[0116] ,

[0117] in: For the actual trajectory Compared with reference trajectory The similarity is in the range of (0,1). It indicates the actual trajectory of changes in the patient's tongue appearance; Indicates the standard reference trajectory; This represents the trajectory length, i.e., the number of time points compared. and Let be the states of the i-th corresponding time point on the two trajectories, and denoted as points on manifold M; The geodesic distance between them; To adjust the parameters (usually set to 0.1-0.3), the sensitivity of similarity to distance differences is controlled; It is a natural exponential function; This represents the mean over all L time points. High similarity (>0.8) indicates a good treatment response that meets expectations; low similarity (<0.5) indicates a poor treatment effect that may require adjustment of the treatment plan.

[0118] Reference Figure 7 The tongue image comparison module 3 is connected to the tongue image measurement module 2 and is used to receive the quantitative analysis results generated by the tongue image measurement module 2, compare and analyze the tongue image images collected at different times, and generate a report on changes before and after treatment.

[0119] The tongue image comparison module 3 includes a tongue color comparison analysis unit 31, a tongue coating quality comparison analysis unit 32, and a tongue shape comparison analysis unit 33.

[0120] The tongue color comparison and analysis unit 31 receives the quantitative analysis results from the tongue image measurement module 2 and compares the color characteristics of the tongue images collected before and after treatment. The system extracts the color change characteristics of the tongue before and after treatment, calculates the geodesic distance on the color manifold, and quantifies the magnitude of change. For standard region divisions (tip, middle, root, left, and right sides of the tongue), the system calculates the color change index for each region.

[0121] ,

[0122] in: For the region The color change index represents the average magnitude of color change in that area. This represents the i-th standard region on the tongue, such as the tip of the tongue, the middle of the tongue, etc. For the region The number of pixels; Indicates the region Pixels in; and Points before and after treatment The color of is represented by a point on manifold M; This is the distance between geodesic lines; Indicates the region The system sums all pixels in the graph. It then matches the change index with typical patterns of symptom changes, interpreting the clinical significance of the changes. For example, a change in tongue tip color from dark red to light red indicates a significant improvement in the state of heart fire syndrome.

[0123] The tongue coating texture comparison and analysis unit 32 receives the quantitative analysis results from the tongue image measurement module 2, compares the tongue coating texture information collected before and after treatment, and outputs the symptom changes. The system analyzes the changes in tongue coating texture, including characteristics such as thickness, distribution, and texture. The system calculates the changes in tongue coating coverage:

[0124] ,

[0125] in: This value represents the change in tongue coating coverage; a positive value indicates an increase, and a negative value indicates a decrease. and These represent the tongue coating coverage rate before and after treatment, respectively, and are defined as the percentage of the tongue coating coverage area to the total tongue area, with a value range of [0, 100%].

[0126] Changes in tongue coating texture are measured by the Euclidean distance of the texture feature vectors:

[0127] ,

[0128] in: Indicates the degree of change in the texture of the tongue coating; and These are tongue coating texture feature vectors before and after treatment, typically 10-20 dimensional vectors, containing features such as texture roughness, regularity, and contrast. This represents the Euclidean norm and calculates the distance between two vectors. The system integrates Traditional Chinese Medicine (TCM) theory to explain the clinical significance of these changes, such as the tongue coating changing from thick and greasy to thin and white, indicating a significant improvement in the internal dampness and turbidity.

[0129] The tongue morphology comparison and analysis unit 33 receives the quantitative analysis results from the tongue image measurement module 2, analyzes the changes in tongue morphology before and after treatment, and evaluates the treatment effect. The system analyzes changes in tongue morphological parameters, including tongue size, edge teeth marks, and tongue fissures. The system calculates the morphological feature change index.

[0130] ,

[0131] in: An index representing the overall change in tongue morphology; This refers to the number of morphological parameters, typically 5-10. and These represent the i-th morphological parameter (such as tongue width, degree of teeth marks, etc.) before and after treatment, respectively. The parameter values ​​are standardized and range from [0,1]. Represents absolute value; express The average value of changes in several morphological parameters. The system combines these changes to assess treatment efficacy, such as reduction of teeth marks on the tongue and improvement of symptoms of spleen deficiency and dampness.

[0132] Reference Figure 8 The tongue image comprehensive analysis module 4 is connected to the tongue image measurement module 2 and the tongue image comparison module 3. It is used to receive the quantitative analysis results of the tongue image measurement module 2 and the change report of the tongue image comparison module 3, to conduct a multi-dimensional comprehensive assessment of health status, and to generate quantitative scores and treatment suggestions.

[0133] The tongue image comprehensive analysis module 4 includes a multi-dimensional feature fusion unit 41, a health status assessment unit 42, and a treatment plan generation unit 43.

[0134] The multi-dimensional feature fusion unit 41 receives the quantitative analysis results from the tongue image measurement module 2 and the change report from the tongue image comparison module 3, integrating the analysis results of tongue color, tongue coating texture, and tongue morphology. The system employs a weighted fusion strategy to integrate multi-dimensional features:

[0135] ,

[0136] in: A comprehensive feature vector is typically a 50-100 dimensional vector that includes features such as color, texture, and shape. For the i-th type of feature vector (such as color feature vector, tongue coating feature vector, morphology feature vector, etc.), the dimensions of each feature vector may be different, and they need to be standardized. This represents the number of feature categories, typically 3-5. Let be the weight coefficient of the i-th type of feature, reflecting the importance of this type of feature in diagnosis, and satisfying the following conditions: In practice, the weighting coefficients are dynamically adjusted according to different syndrome types. For example, for damp-heat syndrome, the weight of tongue coating characteristics is relatively high (about 0.4-0.5); for qi and blood deficiency, the weight of tongue color is relatively high (about 0.4-0.5).

[0137] Based on the integrated analysis results, health status assessment unit 42 calculates a quantitative health status score according to a preset scoring standard. The system establishes a health status assessment model, mapping comprehensive characteristics to a health score of 0-100.

[0138] ,

[0139] in: A quantitative score is given for health status, with a value range of [0, 100]. The higher the score, the better the health status. Let j be the health deviation index, representing the degree to which the tongue appearance characteristics deviate from the health standard, with a value range of [0,1]. The number of deviations from the indicator is usually 5-10; Let be the weight of the j-th deviation indicator, reflecting the importance of that indicator to the health assessment, and satisfying the following conditions: The system ensures that the score is 0 when the deviation is at its maximum. The score is divided into different intervals, corresponding to different health states. For example, 90-100 points indicates a good health state, 70-89 points indicates a slight deviation from the health state, 50-69 points indicates a moderate deviation from the health state, and below 50 points indicates a severe deviation from the health state.

[0140] The treatment plan generation unit 43 generates personalized treatment suggestions based on health status assessment results and traditional Chinese medicine (TCM) theory. The system generates personalized treatment suggestions based on syndrome differentiation and health scores, combined with TCM treatment principles. These suggestions include treatment methods, recommended prescriptions, and lifestyle guidance. The system employs a hybrid reasoning mechanism based on rules and cases, adhering to the fundamental principles of TCM syndrome differentiation and treatment while also referencing successful experiences from similar cases.

[0141] Reference Figure 9The health consultation module 5 is connected to the tongue image comprehensive analysis module 4. It is used to receive the evaluation results of the tongue image comprehensive analysis module 4, query the TCM knowledge graph, compare the diagnosis results with related symptoms, and generate TCM diagnosis results, disease explanations and etiological differentiation.

[0142] The health consultation module 5 includes a knowledge graph query unit 51, a diagnosis result generation unit 52, a symptom explanation unit 53, and a disease etiology dialectics unit 54.

[0143] The knowledge graph query unit 51 receives the evaluation results from the tongue image comprehensive analysis module 4 and retrieves relevant syndrome information from the traditional Chinese medicine knowledge graph. The system maintains a knowledge graph containing traditional Chinese medicine theoretical knowledge, syndrome descriptions, and treatment plans. The knowledge graph adopts an entity-relationship-entity triple structure, covering entity types such as syndromes, symptoms, etiologies, treatment methods, and prescriptions. Based on the syndrome diagnosis results, the system retrieves relevant knowledge nodes to obtain detailed information such as syndrome descriptions, etiologies, pathogenesis, and treatment principles.

[0144] The diagnostic result generation unit 52 generates TCM diagnostic results based on knowledge graph retrieval results and tongue image analysis results. The system integrates the syndrome information from the tongue image analysis results and the knowledge graph to generate standardized TCM diagnostic results. The diagnostic results include syndrome type (such as spleen and stomach damp-heat, liver stagnation and spleen deficiency, etc.), severity (mild, moderate, severe), and accompanying manifestations (such as concurrent yin deficiency, concurrent blood stasis, etc.).

[0145] The Symptom Explanation Unit 53 provides explanations of symptoms from Traditional Chinese Medicine (TCM) theory based on the diagnostic results. The system extracts relevant TCM theoretical explanations of the symptoms from a knowledge graph based on the diagnostic results, including the essential characteristics of the syndrome, its pathogenesis, and typical manifestations. The explanations use easy-to-understand language while retaining the professionalism of TCM theory, helping users understand the medical meaning of the diagnostic results.

[0146] The Etiology Diagnosis Unit 54 analyzes the causes of illness based on diagnostic results and patient information, and provides suggestions for dialectical treatment. The system combines the patient's basic information (such as age, gender, and past medical history) and lifestyle habits to analyze possible causes leading to the current symptoms. The system analyzes the causes from the perspectives of the six pathogenic factors (wind, cold, summer heat, dampness, dryness, and fire) in Traditional Chinese Medicine, the seven emotions, diet, and work-rest balance, providing targeted lifestyle adjustments.

[0147] Reference Figure 10 The data management module 6 is connected to the tongue image acquisition module 1, tongue image measurement module 2, tongue image comparison module 3, tongue image comprehensive analysis module 4, and health consultation module 5. It is used to store historically acquired tongue image data, analysis results, and diagnostic records, manage patient information, support data retrieval and comparison, and provide the data foundation for system operation.

[0148] Data Management Module 6 provides unified management of various types of data generated by the system, including raw tongue images, analysis results, diagnostic records, and patient information. The system employs a structured data storage scheme, establishing a multi-level index structure of patient-tongue image-diagnosis, supporting multi-dimensional data retrieval and analysis. Data storage adopts a hierarchical strategy, with frequently accessed data (such as recent diagnostic results) stored in the fast storage area, and historical data transferred to the archive storage area. The system implements data encryption and access control to ensure patient privacy and data security. Furthermore, the system supports data export functionality, allowing doctors to export diagnostic results in a standard format for easy interaction with other medical systems.

[0149] In summary, the deep learning-based intelligent image analysis system for TCM tongue diagnosis of this invention achieves precise analysis of tongue color features by constructing a color manifold space using differential geometry theory; establishes a scientific mapping relationship between color features and TCM syndromes through geodesic distance calculation; and realizes quantitative tracking of tongue changes and objective evaluation of treatment effects through dynamic evolution analysis. The system integrates functional modules such as tongue image acquisition, measurement, comparison, comprehensive analysis, and health consultation, forming a complete intelligent tongue diagnosis analysis process, providing technical support for the standardization, scientification, and modernization of TCM tongue diagnosis.

[0150] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of protection of the present invention.

Claims

1. A deep learning-based intelligent image analysis system for traditional Chinese medicine tongue diagnosis, characterized in that, The system includes: The tongue image acquisition module is used to acquire standardized tongue images, evaluate the amount of tongue image information, image quality, tongue image location and clarity, and generate a tongue image acquisition quality score. The tongue image measurement module is connected to the tongue image acquisition module and is used to receive the tongue image acquired by the tongue image acquisition module, construct a color manifold space based on the tongue image, calculate the geodesic distance in the color manifold space, establish the mapping relationship between color features and syndromes, perform dynamic evolution analysis of color features, and generate quantitative analysis results of tongue color and tongue coating distribution. The tongue image comparison module is connected to the tongue image measurement module and is used to receive the quantitative analysis results generated by the tongue image measurement module, compare and analyze tongue images collected at different times, and generate a report on changes before and after treatment. The tongue image comprehensive analysis module is connected to the tongue image measurement module and the tongue image comparison module. It is used to receive the quantitative analysis results of the tongue image measurement module and the change report of the tongue image comparison module, to conduct a multi-dimensional comprehensive assessment of health status, and to generate quantitative scores and treatment suggestions. The health consultation module is connected to the tongue image comprehensive analysis module. It is used to receive the evaluation results of the tongue image comprehensive analysis module, query the traditional Chinese medicine knowledge graph, compare the diagnosis results with related symptoms, and generate traditional Chinese medicine diagnosis results, disease explanations, and etiological differentiation.

2. The system according to claim 1, characterized in that, The tongue image measurement module includes: Color manifold construction unit is used to reconstruct the RGB color space in the tongue image into a color manifold with Riemannian metric, construct the tongue color distribution density function, and calculate the manifold curvature features; The geodesic distance calculation unit is connected to the color manifold construction unit and is used to solve the geodesic equation on the color manifold, construct the geodesic distance matrix, and calculate the intrinsic distance between color feature points and symptom reference points. The syndrome mapping analysis unit is connected to the geodesic distance calculation unit and is used to calculate the syndrome similarity score and generate the syndrome probability distribution based on the geodesic distance matrix. The dynamic evolution tracking unit, connected to the color manifold construction unit and the syndrome mapping analysis unit, is used to analyze the changing patterns of the color manifold over time, predict the evolution trend of syndromes, and evaluate the treatment effect.

3. The system according to claim 2, characterized in that, The color manifold building unit includes: The coordinate system reconstruction module is used to define the local coordinate system of the color manifold, establish mapping functions, and ensure smooth transitions between regions. The metric tensor construction module, connected to the coordinate system reconstruction module, is used to define the Riemann metric tensor on the manifold, optimize the sensitivity of the metric to the syndrome features, and ensure the continuity of the metric. The manifold feature extraction module, connected to the metric tensor construction module, is used to calculate the curvature features of the manifold, identify topological features, and integrate regional features.

4. The system according to claim 2, characterized in that, The geodesic distance calculation unit includes: The mesh structure design module is used to build a mesh structure that adapts to tongue images, mapping pixels to mesh nodes on a manifold; The geodesic equation solving module, connected to the grid structure design module, is used to discretize the continuous geodesic equation, realize adaptive time step control, and solve the geodesic equation. The key point selection module is connected to the geodesic equation solving module and is used to extract representative color points, identify sensitive areas of syndrome, and control sampling density. The distance matrix construction module, connected to the key point selection module and the geodesic equation solving module, is used to calculate the geodesic distance between key points, interpolate the distance values ​​of non-key points, and optimize the storage and retrieval of the distance matrix.

5. The system according to claim 2, characterized in that, The syndrome mapping analysis unit includes: The syndrome feature space construction module is used to define the syndrome feature representation framework, integrate expert knowledge, and optimize syndrome templates; The basic mapping design module, connected to the syndrome feature space construction module, is used to design the mapping function from color manifold features to the syndrome space and learn complex nonlinear mapping relationships. A multi-scale similarity calculation module, connected to the basic mapping design module, is used to calculate the similarity of local regions, aggregate global similarity, and perform context-sensitive adjustments. The syndrome probability distribution generation module is connected to the multi-scale similarity calculation module and is used to convert similarity into probability values, analyze syndrome combination patterns, and quantify diagnostic uncertainty.

6. The system according to claim 2, characterized in that, The dynamic evolution tracking unit includes: The time-series data preprocessing module is used to align tongue image data from multiple time points, normalize timestamps, and handle abnormal data. The thermal equation simulation module, connected to the time-series data preprocessing module, is used to solve the thermal equation on the manifold and analyze the diffusion process of color features. The curvature flow analysis module, connected to the time-series data preprocessing module, is used to simulate the evolution of manifold curvature flow, track the trajectory of feature points, and predict the steady state. The treatment effect evaluation module, connected to the thermal equation simulation module and the curvature flow analysis module, is used to identify treatment response patterns, build an evolution path library, compare actual trajectories with reference paths, and generate treatment suggestions.

7. The system according to claim 1, characterized in that, The tongue image comparison module includes: The tongue color comparison and analysis unit is used to receive the quantitative analysis results of the tongue image measurement module and compare and analyze the color features of the tongue image photos collected before and after treatment. The tongue coating quality comparison and analysis unit is used to receive the quantitative analysis results of the tongue image measurement module, compare the tongue coating quality information collected before and after treatment, and output the symptom changes. The tongue morphology comparison and analysis unit is used to receive the quantitative analysis results of the tongue image measurement module, analyze the changes in tongue morphology before and after treatment, and evaluate the treatment effect.

8. The system according to claim 1, characterized in that, The tongue image comprehensive analysis module includes: The multi-dimensional feature fusion unit is used to receive the quantitative analysis results of the tongue image measurement module and the change report of the tongue image comparison module, and integrate the analysis results of tongue color, tongue coating and tongue morphology; A health status assessment unit, connected to the multidimensional feature fusion unit, is used to calculate a quantitative health status score based on the integrated analysis results and according to a preset scoring standard. The treatment plan generation unit is connected to the health status assessment unit and is used to generate personalized treatment suggestions based on the health status assessment results and traditional Chinese medicine theory.

9. The system according to claim 1, characterized in that, The health consultation module includes: The knowledge graph query unit is used to receive the evaluation results of the tongue image comprehensive analysis module and retrieve relevant syndrome information in the traditional Chinese medicine knowledge graph; The diagnosis result generation unit is connected to the knowledge graph query unit and is used to generate traditional Chinese medicine diagnosis results based on the knowledge graph retrieval results and tongue image analysis results. The symptom interpretation unit, connected to the diagnosis result generation unit, is used to provide symptom interpretations in traditional Chinese medicine theory based on the diagnosis results. The etiology dialectics unit, connected to the diagnosis result generation unit, is used to analyze the etiology based on the diagnosis result and patient information, and provide dialectical treatment suggestions.

10. The system according to claim 1, characterized in that, The system also includes a data management module, which is connected to the tongue image acquisition module, the tongue image measurement module, the tongue image comparison module, the tongue image comprehensive analysis module, and the health consultation module. The data management module is used to store historically acquired tongue image data, analysis results, and diagnostic records, manage patient information, support data retrieval and comparison, and provide the data foundation for system operation.

Images