A tongue appearance traditional chinese medicine zangfu information prediction analysis method and system

Abstract

The present application belongs to the technical field of image analysis, and relates to a tongue image traditional Chinese medicine zangfu information prediction and analysis method and system. By extracting a plurality of local features corresponding to zangfu reflection areas in a deep visual feature tensor to construct an initial node feature matrix with zangfu as nodes, the model focuses on the zangfu corresponding areas with pathological significance in the tongue image by fusing traditional Chinese medicine zangfu anatomy prior knowledge, irrelevant background noise interference is avoided, a static prior adjacency matrix is constructed by presetting zangfu pathological transmission prior, and a dynamic adjacency matrix is generated by combining a learnable parameter matrix, so that the graph structure can adapt to the traditional Chinese medicine zangfu transmission law and self-adaptively adjust the pathological correlation weight between zangfu, the accuracy is improved in the traditional Chinese medicine syndrome type classification prediction of tongue diagnosis, especially in the traditional Chinese medicine syndrome type classification prediction of psoriasis, the method disclosed in the present application introduces the explicit explainability of traditional Chinese medicine zangfu theory in the reasoning process, and improves the prediction accuracy.

Description

Technical Field

[0001] This invention belongs to the field of image analysis technology and relates to a method and system for predicting and analyzing information about the internal organs in traditional Chinese medicine based on tongue images. Background Technology

[0002] Tongue diagnosis in Traditional Chinese Medicine (TCM) is a core technique of TCM observation and a key indicator for the objectification and visualization of TCM clinical diagnosis and treatment. Based on the TCM theory of tongue-visceral correspondence and the modern medical oral mucosa-visceral reflex mechanism, different anatomical regions of the tongue have precise correspondences with the functions and pathological states of the body's internal organs. Specifically, the tip of the tongue corresponds to the heart and lungs, the middle to the spleen and stomach, the root to the kidneys, and the sides to the liver and gallbladder. This pattern provides a stable basis for non-invasive auxiliary diagnosis of visceral diseases. In the clinical diagnosis and treatment of chronic inflammatory skin diseases such as psoriasis, tongue appearance is the gold standard for distinguishing the three core types of blood-heat syndrome, blood-stasis syndrome, and blood-dryness syndrome. It directly guides the selection of TCM prescriptions, external treatment plans, and prognosis. Furthermore, changes in tongue appearance can objectively reflect the patient's skin inflammation level, systemic immune dysregulation, peripheral microcirculatory disorders, and oral mucosal involvement. Geographic tongue and fissured tongue are typical oral manifestations of psoriasis, possessing significant biomedical diagnostic and disease monitoring value.

[0003] While existing deep learning-based automated tongue diagnosis technologies have made some progress, they still face key shortcomings in real-world clinical translation and medical application. First, current models are primarily data-driven, heavily reliant on large-scale, high-quality expert-annotated data. This makes them prone to overfitting in clinical settings with small sample sizes and low annotation rates, resulting in inadequate generalization and diagnostic robustness. Second, conventional deep learning methods treat the tongue image merely as a simple pixel image, completely ignoring the inherent holographic theory of internal organs and anatomical regions in traditional Chinese medicine (TCM) tongue diagnosis. Feature extraction is easily affected by noise from the oral cavity background, lighting, and tongue posture, leading to a lack of medical rationality in the diagnostic results. Finally, current technologies cannot explicitly model and interpretably reason about the potential pathological connections in TCM's concept of "interconnection between internal organs and the relationship between exterior and interior," resulting in a "black box" diagnostic process that is difficult for clinicians to accept and recognize, severely limiting the clinical credibility and widespread application of intelligent tongue diagnosis in TCM. Therefore, a highly accurate, robust, and interpretable intelligent tongue diagnosis method is still needed in clinical practice to address these shortcomings. Summary of the Invention

[0004] The purpose of this invention is to address the problems in existing technologies, such as the scarcity of labeled samples in clinical tongue diagnosis scenarios, the neglect of prior knowledge of TCM visceral anatomy in existing deep learning methods, and the lack of interpretability, which lead to low diagnostic accuracy, poor robustness, and difficulty in gaining the trust of clinicians. This invention provides a method and system for predicting and analyzing TCM visceral information from tongue images.

[0005] To achieve the above objectives, the present invention employs the following technical solution: A method for predicting and analyzing the information of internal organs in traditional Chinese medicine based on tongue appearance includes the following steps: Obtain standardized tongue images; Extract deep visual feature tensors from standardized tongue images; Local features corresponding to multiple visceral reflex zones are extracted from the deep visual feature tensor, and an initial node feature matrix based on visceral nodes is constructed. A static prior adjacency matrix is ​​constructed based on the pre-defined pathological transmission prior between organs. A learnable parameter matrix with the same dimension as the static prior adjacency matrix is ​​constructed. A dynamic adjacency matrix is ​​generated based on the static prior adjacency matrix and the learnable parameter matrix. The initial node feature matrix and the dynamic adjacency matrix are subjected to graph structure information propagation and feature update to obtain the updated node feature matrix. Based on the updated node feature matrix, classification and prediction information for TCM syndrome types of psoriasis is generated.

[0006] A further improvement of the present invention is that: The acquisition of standardized tongue images includes: The input image is processed by an object detector to generate a bounding box cue for the tongue region; The bounding box prompts are used to segment the large model and generate a semantic segmentation mask; The pure tongue body is extracted based on semantic segmentation mask, and then the extracted pure tongue body is scaled to a fixed resolution through geometric transformation to obtain a standardized tongue body image.

[0007] The extraction of deep visual feature tensors from standardized tongue images includes: Standardized tongue images are input into a ResNet-18 backbone network to extract deep visual feature tensors from the tongue images. .

[0008] The step of extracting local features corresponding to multiple visceral reflex zones from the deep visual feature tensor and constructing an initial node feature matrix of visceral nodes includes: Based on the anatomical coordinates of the internal organs in traditional Chinese medicine, the key anatomical coordinates corresponding to the reflex zones of the five internal organs—heart, liver, spleen, lung, and kidney—are determined on the deep visual feature tensor. Regional average pooling is performed on the predetermined radius region around each key anatomical coordinate to obtain the initial node embedding vector of the corresponding visceral node; The initial node embedding vectors of all organ nodes are concatenated to obtain the initial node feature matrix.

[0009] The generation of a dynamic adjacency matrix based on a static prior adjacency matrix and a learnable parameter matrix includes: Construct a binary static star topology matrix The binarized static star topology matrix is ​​used to simulate the physiological and pathological relationships between the internal organs; The static prior adjacency matrix is ​​added to the learnable parameter matrix, and a non-negativity constraint is applied using the ReLU nonlinear activation function to generate the dynamic adjacency matrix.

[0010] in, Represents the learnable parameter matrix; This represents the static prior adjacency matrix.

[0011] The step of performing graph structure information propagation and feature updating on the initial node feature matrix and the dynamic adjacency matrix to obtain the updated node feature matrix includes: The dynamic adjacency matrix is ​​input into the graph convolutional network for inter-layer propagation. The propagation rules of the layers are as follows:

[0012] in, It is the first The weight matrix of the layer, It is a non-linear activation function.

[0013] This also includes training the model using a semi-supervised training method: The network parameters are trained by jointly using labeled and unlabeled samples, wherein the number of labeled samples is less than the number of unlabeled samples. After performing weak data augmentation on labeled samples, the cross-entropy loss is calculated, and the calculation result is used as the supervision loss. Based on unlabeled samples, a predicted distribution is obtained through weak data augmentation and pseudo-labels are generated. When the maximum confidence of the weakly augmented prediction is greater than a preset threshold, strong data augmentation is performed on the corresponding unlabeled sample. The cross-entropy loss between the augmented predicted distribution and the pseudo-label is calculated and used as the consistency loss. The total loss is obtained based on the supervision loss and consistency loss, and the network parameters are jointly updated using the total loss.

[0014] A tongue-image-based traditional Chinese medicine (TCM) organ information prediction and analysis system includes: The raw image processing module is used to acquire standardized tongue images; The feature extraction module is used to extract the deep visual feature tensor of the standardized tongue image; The initial node feature matrix construction module is used to extract local features corresponding to multiple visceral reflex zones from the deep visual feature tensor and construct an initial node feature matrix based on visceral nodes. The dynamic adjacency matrix construction module is used to construct a static prior adjacency matrix based on a preset prior of pathological transmission between organs, construct a learnable parameter matrix with the same dimension as the static prior adjacency matrix, and generate a dynamic adjacency matrix based on the static prior adjacency matrix and the learnable parameter matrix. The prediction module is used to propagate graph structure information and update features between the initial node feature matrix and the dynamic adjacency matrix to obtain the updated node feature matrix. Based on the updated node feature matrix, classification prediction information for TCM syndrome types of psoriasis is generated.

[0015] A terminal device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of any of the methods described above.

[0016] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the methods described herein.

[0017] Compared with the prior art, the present invention has the following beneficial effects: This invention discloses a method for predicting and analyzing the information of internal organs in traditional Chinese medicine (TCM) based on tongue appearance. It constructs an initial node feature matrix with internal organs as nodes by extracting local features corresponding to multiple organ reflex zones from a deep visual feature tensor. This matrix integrates prior knowledge of TCM organ anatomy, allowing the model to focus on the pathologically significant organ-corresponding regions in the tongue appearance, avoiding interference from irrelevant background noise. Simultaneously, it constructs a static prior adjacency matrix by pre-setting prior knowledge of pathological transmission between organs and generates a dynamic adjacency matrix by combining it with a learnable parameter matrix. This ensures the graph structure adapts to the transmission patterns of TCM organs and can adaptively adjust the pathological association weights between organs, improving accuracy in predicting TCM syndrome types in tongue diagnosis. Particularly for predicting TCM syndrome types in psoriasis, the method disclosed in this invention introduces explicit interpretability of TCM organ theory during the reasoning process, improving the accuracy of prediction. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the overall process framework disclosed in the embodiments of the present invention; Figure 2The above are schematic diagrams of tongue segmentation using the TongueSAM method disclosed in this embodiment of the invention, showing the segmentation effects of blood heat syndrome, blood stasis syndrome, and blood dryness syndrome, in sequence. Figure 3 This is a diagram showing the corresponding structural regions of the tongue surface and internal organs as disclosed in an embodiment of the present invention. Figure 4 This is a visualization diagram of a high-confidence prediction sample disclosed in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0021] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0022] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0023] In the description of the embodiments of the present invention, it should be noted that if terms such as "upper," "lower," "horizontal," or "inner" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of the invention is in use, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. Furthermore, terms such as "first" and "second" are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0024] Furthermore, the use of the term "horizontal" does not imply that the component must be absolutely horizontal, but rather that it can be slightly tilted. For example, "horizontal" simply means that its direction is more horizontal than "vertical," and does not mean that the structure must be completely horizontal, but can be slightly tilted.

[0025] In the description of the embodiments of the present invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention according to the specific circumstances.

[0026] The present invention will now be described in further detail with reference to the accompanying drawings: See Figure 1 This invention discloses a method for predicting and analyzing information about the internal organs in traditional Chinese medicine based on tongue images. Specifically, it is a semi-supervised method for diagnosing diseases based on tongue images that integrates prior knowledge of the anatomy of the internal organs in traditional Chinese medicine with an adaptive graph neural network. This embodiment is mainly proposed to meet the clinical auxiliary diagnostic needs of chronic diseases such as psoriasis. It realizes a fully intelligent diagnosis process that includes standardized tongue image acquisition, extraction of medical anatomical features, reasoning of the internal organ association graph, and semi-supervised disease classification, significantly improving medical reliability, clinical interpretability, and generalization ability for small samples.

[0027] The main steps include: Step 1: Tongue image standardization preprocessing based on a large medical segmentation model To eliminate interference from non-target areas such as the oral cavity background, lips, and teeth, and to ensure the consistency of clinical images, this invention constructs a dedicated medical segmentation pipeline for tongue images.

[0028] Let the original input image space be First, an object detector (YOLOX) is used as the cue generator. For the input image Generate bounding box hints for the tongue region. :

[0029] Furthermore, an image encoder utilizing the Segment Anything Model (SAM) is employed. and prompt encoder Generate semantic segmentation mask :

[0030] like Figure 2 As shown, the final standardized input image Through the Hadamard product (symbol: Extract the pure tongue body using a mask and then perform geometric transformations. Interpolate the image to a fixed resolution. To ensure spatial consistency in subsequent feature extraction:

[0031] Step 2: Extraction of medical features based on visceral anatomical coordinates The deep visual feature tensor of the tongue image was extracted using a ResNet-18 backbone network. ,in ( For the number of channels, and (The height and width of the feature map) correspond to clinical and pathological features such as tongue color, tongue coating thickness, texture, moisture, petechiae and ecchymosis, and geographic tongue atrophy.

[0032] like Figure 3 As shown, based on the atlas of tongue diagnosis in Traditional Chinese Medicine, a set of key anatomical coordinates is defined on the two-dimensional plane of the feature map. These correspond to the five organ reflex zones: heart (tip of the tongue), liver (right side of the tongue), spleen (middle of the tongue), lungs (left side of the tongue), and kidneys (root of the tongue). See details... Figure 2 .

[0033] To enhance robustness to non-rigid tongue deformations (such as differences in patient tongue protrusion posture), this embodiment of the invention abandons single-pixel sampling and employs a region-average pooling operator. For the first Each anatomical node has a node embedding vector. Defined as:

[0034] in, For the set pooling radius, in the examples of this invention ,correspond The receptive field neighborhood is determined. The vectors obtained from pooling the five regions are concatenated to construct the initial node feature matrix of the graph neural network. Each node corresponds to a pathological state characteristic of an organ.

[0035] Step 3: Adaptive dynamic graph reasoning that integrates TCM prior knowledge First, based on the TCM pathological transmission pattern of "spleen and stomach as the center, and the five internal organs working together" as the initial prior, a binary static star-shaped topological matrix is ​​constructed. It simulates the physiological and pathological connections between the internal organs. Specifically, the blood-heat syndrome of psoriasis is commonly characterized by: excessive fire in the heart and liver, and damp-heat in the spleen and stomach.

[0036] Furthermore, in order to overcome the limitations of static priors and adapt to specific diseases, this embodiment of the invention introduces a learnable parameter matrix with the same dimension as the prior matrix. Furthermore, a nonnegativity constraint is imposed by applying a nonlinear activation function (ReLU) to generate a dynamic adjacency matrix. :

[0037] This mechanism allows the model to adaptively enhance or suppress pathological dependency weights between specific organs during backpropagation. If a weight update is less than 0, ReLU truncates it to 0, effectively breaking the edge connection. This matrix can adaptively strengthen / weaken pathological connections between organs, such as automatically increasing the weight of "heart → liver" in blood-heat syndrome and the weight of "spleen → kidney" in blood stasis syndrome, thus enabling the model to have clinical and pathological interpretability.

[0038] Furthermore, the feature matrix is ​​input into the graph convolutional network (GCN) for inter-layer propagation, the 1st... The propagation rules of the layers are formalized as follows:

[0039] in, It is the first The weight matrix of the layer, It is a non-linear activation function. After multiple layers of graph convolution, the final psoriasis type prediction result is output through a fully connected layer, such as: psoriasis due to blood heat, blood stasis, or blood dryness.

[0040] Step 4: Semi-supervised joint training To address the scarcity of clinical labeled data, a semi-supervised training strategy combining weak supervision and strong consistency regularization is employed, specifically including: Suppose that there is a very small number of labeled datasets. Massive unlabeled datasets for Total loss function To monitor losses And unsupervised consistency loss Linear combination:

[0041] Supervised items (labeled data): Weak data augmentation of labeled samples. After (e.g., rotation, translation), calculate the standard cross-entropy loss:

[0042] Unsupervised terms (unlabeled data): A consistency constraint strategy is adopted for unlabeled samples. First, the distribution is predicted using weak enhancement calculations. And generate pseudo tags. Only when the maximum confidence level of the weakly enhanced prediction is greater than a preset threshold. In this experiment, strong data augmentation with drastic perturbations was only applied to the sample when the set value was 0.95. And calculate its consistency loss with pseudo-labels:

[0043] in, This is the indicator function. Finally, by calculating the total gradient, the parameters of the backbone network, graph convolutional layers, and dynamic matrices in the system are jointly updated. All weights.

[0044] Even under extreme clinical conditions with an annotation rate of only 11.2% (90 annotated images), the embodiments of the present invention can still achieve stable training, avoid overfitting, and reach the diagnostic accuracy of a traditional Chinese medicine physician.

[0045] Step 5: Clinical Applications and Downstream Task Expansion Intelligent tongue diagnosis, as a key technology for the objectification of traditional Chinese medicine (TCM) and the auxiliary diagnosis and treatment of chronic diseases, can provide stable support for multiple downstream clinical tasks. In the diagnosis and treatment of chronic skin diseases such as psoriasis, it can quickly identify TCM syndromes such as blood heat, blood stasis, and blood dryness, providing objective evidence for individualized prescriptions and improving the consistency of diagnosis and efficacy. In chronic disease screening and follow-up, it can non-invasively and conveniently complete the initial screening of high-risk groups and long-term disease monitoring, reducing invasive examinations and lowering outpatient and follow-up costs. In TCM telemedicine and primary healthcare scenarios, it can provide standardized and interpretable tongue diagnosis conclusions, compensating for the lack of experience of primary care physicians and improving the efficiency of hierarchical diagnosis and treatment. In addition, this method can help analyze the correlation mechanism between tongue appearance and organ function and pathological changes, providing objective evidence for TCM diagnosis and treatment through visualized organ correlation weights, and promoting the standardization and clinical implementation of precision TCM and intelligent TCM.

[0046] Step 6: Experimental Verification and Effect Evaluation This invention uses a clinical tongue image dataset of psoriasis for verification experiments. The original dataset contains 900 clinical tongue images. After manual screening to remove unqualified samples such as blurry images, incomplete tongue exposure, and lip interference, an experimental dataset containing 803 high-quality psoriasis tongue images was finally constructed. All samples were obtained in a standardized TCM clinical collection environment and were uniformly labeled by senior TCM physicians according to TCM diagnosis and treatment standards as three typical types: blood heat syndrome, blood stasis syndrome, and blood dryness syndrome.

[0047] To meet the actual needs of scarce labeled samples and small-sample learning in real clinical scenarios, this experiment adopted a data partitioning method with an extremely low labeling rate. 90 images were randomly selected (30 for each type) as the labeled training set, and the remaining 713 images were used as the unlabeled training set. The overall labeling rate was only 11.2%, which was used to fully verify the diagnostic ability and generalization performance of the embodiments of the present invention under extremely low sample conditions.

[0048] The model was trained using a semi-supervised training configuration based on FixMatc, with stochastic gradient descent (SGD) as the optimizer, momentum set to 0.9, and weight decay coefficient of 5 × 10⁻⁶. -4 The batch size was set to 8 for labeled data and 56 for unlabeled data. The confidence threshold for pseudo-label screening was set to 0.95 to ensure that high-confidence samples participated in model optimization. The initial learning rate was 0.003, and a cosine annealing strategy was used to gradually decay to 0 within 50 iterations to ensure a stable and reliable training process.

[0049] Training dynamics and convergence results demonstrate that the method of this invention achieves rapid and stable convergence even with extremely low labeled sample conditions, efficiently learning and extracting discriminative visceral anatomical and pathological features from a very small amount of labeled data. Simultaneously, unlabeled data can continuously provide effective gradients to the model through consistency regularization, significantly suppressing overfitting and model degradation, enabling the system to maintain excellent stability and generalization ability even in small-sample clinical scenarios.

[0050] like Figure 4 As shown, this invention can accurately identify the typical tongue pathological features of different syndromes of psoriasis. For blood-heat syndrome, the model can accurately capture key features such as a red tongue, pinpoints on the tongue surface, and congestion at the tip and sides of the tongue, which is highly consistent with the organ reflex zones of "heart and liver fire" in traditional Chinese medicine. For blood stasis syndrome, the model can effectively distinguish features such as pathological purplish-dark tongue, ecchymosis, and petechiae, and has a strong ability to perceive local texture abnormalities. For blood-dryness syndrome, the model can stably identify manifestations such as dry tongue surface, lack of saliva, and cracks, and has high robustness to different types of cracks. Confidence analysis shows that the model's high-confidence prediction sample scores are all above 0.999, with some reaching 1.0000, indicating that based on anatomical prior guidance and TongueSAM standardized segmentation, the model can form a clear and separable decision boundary in the feature space, achieving expert-level diagnostic accuracy with extremely low annotation rates. At the same time, the model's reasoning process is highly consistent with the theory of organ classification in traditional Chinese medicine, and the diagnostic process is transparent and interpretable, fully meeting the clinical needs for the reliability and interpretability of intelligent auxiliary diagnosis in traditional Chinese medicine.

[0051] As can be seen from the above technical solution, the semi-supervised tongue diagnosis method that integrates anatomical priors and adaptive graph neural networks disclosed in this invention can efficiently decode the pathological manifestations of viscera and organs in traditional Chinese medicine from clinical tongue images. First, based on the theory of visceral classification in tongue diagnosis and the anatomical characteristics of the oral mucosa, the tongue image region is transformed into a structured visceral node graph. Combining the advantages of large-scale medical segmentation models and dynamic graph reasoning, an anatomical prior-guided tongue image feature extraction module is constructed to accurately remove background noise and effectively mine the local pathological features of the tongue image and the spatial correlation information of viscera and organs. Then, through preset anatomical coordinates and adaptive adjacency matrix learning, the pathological transmission relationship between viscera and organs is modeled to simulate the real diagnostic logic of TCM clinical syndrome differentiation. Finally, through semi-supervised consistency constraints and graph convolution feature fusion, the pathological changes in the tongue surface region, the functional correlation between viscera and organs, and the correspondence between tongue image and disease type are accurately captured, significantly improving the accuracy, robustness, and clinical interpretability of tongue image syndrome differentiation diagnosis.

[0052] This embodiment also discloses a tongue image-based traditional Chinese medicine organ information prediction and analysis system, including: The raw image processing module is used to acquire standardized tongue images; The feature extraction module is used to extract the deep visual feature tensor of the standardized tongue image; The initial node feature matrix construction module is used to extract local features corresponding to multiple visceral reflex zones from the deep visual feature tensor and construct an initial node feature matrix based on visceral nodes. The dynamic adjacency matrix construction module is used to construct a static prior adjacency matrix based on a preset prior of pathological transmission between organs, construct a learnable parameter matrix with the same dimension as the static prior adjacency matrix, and generate a dynamic adjacency matrix based on the static prior adjacency matrix and the learnable parameter matrix. The prediction module is used to propagate graph structure information and update features between the initial node feature matrix and the dynamic adjacency matrix to obtain the updated node feature matrix. Based on the updated node feature matrix, classification prediction information for TCM syndrome types of psoriasis is generated.

[0053] This invention utilizes TongueSAM to achieve standardized clinical preprocessing and zero-sample precise tongue segmentation, effectively eliminating oral noise interference and demonstrating high robustness to complex tongue appearances in psoriasis, such as geographic tongue and fissured tongue. It pioneers an visceral anatomical coordinate anchoring and dynamic graph reasoning mechanism, transforming the TCM pathological pattern of "five viscera linkage" into a learnable topological structure, enabling the model to possess clear pathological interpretability, solving the black-box problem of deep learning and aligning with the diagnostic logic of clinicians. Employing semi-supervised consistency constraint training, it achieves stable learning with minimal labeled data, significantly reducing clinical labeling costs and making it suitable for real-world applications such as primary care hospitals and remote tongue diagnosis. Furthermore, it possesses strong disease specificity, accurately adapting to typical pathological tongue features in psoriasis, such as red tongue, ecchymosis, peeling tongue coating, and fissured tongue, and can simultaneously complete TCM syndrome differentiation and disease severity assessment, providing reliable support for precise TCM diagnosis and treatment of psoriasis.

[0054] A schematic diagram of a terminal device according to an embodiment of the present invention. The terminal device of this embodiment includes: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps in the various method embodiments described above. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the various device embodiments described above.

[0055] The computer program can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention.

[0056] The terminal device can be a desktop computer, laptop computer, cloud server, or other device with strong computing power. The terminal device may include, but is not limited to, a processor and memory.

[0057] The optimal choice for the processor is a multi-core high-speed central processing unit (CPU).

[0058] The memory can be used to store the computer program and / or module. The processor implements various functions of the terminal device by running or executing the computer program and / or module stored in the memory and calling the data stored in the memory.

[0059] If the modules / units integrated into the terminal device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0060] The above are merely preferred embodiments of the present invention and are not intended to limit the present 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 principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting and analyzing information about internal organs in traditional Chinese medicine based on tongue appearance, characterized in that, Includes the following steps: Obtain standardized tongue images; Extract deep visual feature tensors from standardized tongue images; Local features corresponding to multiple visceral reflex zones are extracted from the deep visual feature tensor, and an initial node feature matrix based on visceral nodes is constructed. A static prior adjacency matrix is ​​constructed based on the pre-defined pathological transmission prior between organs. A learnable parameter matrix with the same dimension as the static prior adjacency matrix is ​​constructed. A dynamic adjacency matrix is ​​generated based on the static prior adjacency matrix and the learnable parameter matrix. The initial node feature matrix and the dynamic adjacency matrix are subjected to graph structure information propagation and feature update to obtain the updated node feature matrix. Based on the updated node feature matrix, classification and prediction information for TCM syndrome types of psoriasis is generated.

2. The method for predicting and analyzing the information of internal organs in traditional Chinese medicine based on tongue appearance according to claim 1, characterized in that, The acquisition of standardized tongue images includes: The input image is processed by an object detector to generate bounding box prompts for the tongue region; The bounding box prompts are used to segment the large model and generate a semantic segmentation mask; The pure tongue body is extracted based on semantic segmentation mask, and then the extracted pure tongue body is scaled to a fixed resolution through geometric transformation to obtain a standardized tongue body image.

3. The method for predicting and analyzing information about internal organs in traditional Chinese medicine based on tongue appearance according to claim 1, characterized in that, The extraction of deep visual feature tensors from standardized tongue images includes: Standardized tongue images are input into a ResNet-18 backbone network to extract deep visual feature tensors from the tongue images. .

4. The method for predicting and analyzing the information of internal organs in traditional Chinese medicine based on tongue appearance according to claim 1, characterized in that, The step of extracting local features corresponding to multiple visceral reflex zones from the deep visual feature tensor and constructing an initial node feature matrix of visceral nodes includes: Based on the anatomical coordinates of the internal organs in traditional Chinese medicine, the key anatomical coordinates corresponding to the reflex zones of the five internal organs—heart, liver, spleen, lung, and kidney—are determined on the deep visual feature tensor. Regional average pooling is performed on the predetermined radius region around each key anatomical coordinate to obtain the initial node embedding vector of the corresponding visceral node; The initial node embedding vectors of all organ nodes are concatenated to obtain the initial node feature matrix.

5. The method for predicting and analyzing the information of internal organs in traditional Chinese medicine based on tongue appearance according to claim 1, characterized in that, The generation of a dynamic adjacency matrix based on a static prior adjacency matrix and a learnable parameter matrix includes: Construct a binary static star topology matrix The binarized static star topology matrix is ​​used to simulate the physiological and pathological relationships between the internal organs; The static prior adjacency matrix is ​​added to the learnable parameter matrix, and a non-negativity constraint is applied using the ReLU nonlinear activation function to generate the dynamic adjacency matrix. in, Represents the learnable parameter matrix; This represents the static prior adjacency matrix.

6. The method for predicting and analyzing the information of internal organs in traditional Chinese medicine based on tongue appearance according to claim 1, characterized in that, The step of performing graph structure information propagation and feature updating on the initial node feature matrix and the dynamic adjacency matrix to obtain the updated node feature matrix includes: The dynamic adjacency matrix is ​​input into the graph convolutional network for inter-layer propagation. The propagation rules of the layers are as follows: in, It is the first The weight matrix of the layer, It is a non-linear activation function.

7. The method for predicting and analyzing the information of internal organs in traditional Chinese medicine based on tongue appearance according to claim 1, characterized in that, This also includes training the model using a semi-supervised training method: The network parameters are trained by jointly using labeled and unlabeled samples, wherein the number of labeled samples is less than the number of unlabeled samples. After performing weak data augmentation on labeled samples, the cross-entropy loss is calculated, and the calculation result is used as the supervision loss. Based on unlabeled samples, a predicted distribution is obtained through weak data augmentation and pseudo-labels are generated. When the maximum confidence of the weakly augmented prediction is greater than a preset threshold, strong data augmentation is performed on the corresponding unlabeled sample. The cross-entropy loss between the augmented predicted distribution and the pseudo-label is calculated and used as the consistency loss. The total loss is obtained based on the supervision loss and consistency loss, and the network parameters are jointly updated using the total loss.

8. A tongue-image-based traditional Chinese medicine organ information prediction and analysis system, characterized in that, include: The raw image processing module is used to acquire standardized tongue images; The feature extraction module is used to extract the deep visual feature tensor of the standardized tongue image; The initial node feature matrix construction module is used to extract local features corresponding to multiple visceral reflex zones from the deep visual feature tensor and construct an initial node feature matrix based on visceral nodes. The dynamic adjacency matrix construction module is used to construct a static prior adjacency matrix based on a preset prior of pathological transmission between organs, construct a learnable parameter matrix with the same dimension as the static prior adjacency matrix, and generate a dynamic adjacency matrix based on the static prior adjacency matrix and the learnable parameter matrix. The prediction module is used to propagate graph structure information and update features between the initial node feature matrix and the dynamic adjacency matrix to obtain the updated node feature matrix. Based on the updated node feature matrix, classification prediction information for TCM syndrome types of psoriasis is generated.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-7.

Images