A rheumatoid arthritis tongue diagnosis recognition method, device and medium based on information fusion

Abstract

The application provides a rheumatoid arthritis tongue diagnosis recognition method and device based on information fusion and a medium, and belongs to the tongue image recognition technical field. The method comprises the following steps: extracting a tongue image area from an original tongue image; segmenting the tongue image area by using an improved AttentionU-Net network; identifying a tongue fur area from a tongue body image by using a trained tongue body segmentation model; training the tongue body segmentation model by using a transfer learning strategy, and performing feature selection and sample balance processing on training set data; and establishing a rheumatoid arthritis tongue diagnosis recognition model based on a self-adaptive Stacking strategy of prediction entropy. In the rheumatoid arthritis tongue diagnosis recognition model, the prediction entropy of a base learner is used as the weight of the base learner, the output probability of the base learner is used to generate weighted feature input, and a meta-learner outputs a prediction result. The application solves the problem of low detection accuracy caused by unbalanced tongue image samples in the prior art.

Description

Technical Field

[0001] This invention relates to a method, device, and medium for tongue diagnosis in rheumatoid arthritis based on information fusion, belonging to the field of tongue image recognition technology. Background Technology

[0002] Rheumatoid arthritis (RA) is a systemic autoimmune disease of unknown etiology, characterized by synovitis and joint pain, ultimately eroding cartilage and destroying bone. It can also cause damage to multiple systems, including the heart, lungs, and kidneys, leading to permanent disability and shortened life expectancy in severe cases. Therefore, improving the early recognition rate of RA is crucial for controlling disease progression, improving prognosis, and enhancing quality of life in RA patients.

[0003] By segmenting the tongue image to remove redundant background information such as facial areas from the original tongue image, and using computer technology and image processing methods to analyze the tongue image image and extract tongue image features, we can quantitatively and objectively assist in the identification of diseases and reduce the errors caused by doctors' personal experience.

[0004] Current research on the objectification of tongue diagnosis mainly focuses on the standardization and quantitative analysis of tongue appearance features. There are few specific studies summarizing the objectification of tongue diagnosis for rheumatoid arthritis. Moreover, these studies focus on the correlation between tongue appearance features and rheumatoid arthritis, and most of them use statistical analysis methods. Existing literature rarely shows research on constructing a rheumatoid arthritis recognition model based on tongue appearance using artificial intelligence technology and image processing methods.

[0005] Furthermore, tongue image datasets often suffer from class imbalance. During the training of tongue diagnosis models, this imbalance can cause numerous problems for disease identification and classification, leading to training results biased towards the majority class and poorer fitting of minority class samples, resulting in lower classification accuracy. When using tongue diagnosis models for disease screening, the number of normal tongue image samples is far greater than that of patient tongue image samples. If the data imbalance problem is not considered, patient tongue image samples will underfit, leading to a higher false negative rate and biased identification results.

[0006] Therefore, existing methods for tongue diagnosis in rheumatoid arthritis cannot achieve rapid and accurate detection of tongue appearance. Summary of the Invention

[0007] The purpose of this invention is to provide a method, device, and medium for tongue diagnosis recognition in rheumatoid arthritis based on information fusion, which solves the problem of low detection accuracy caused by imbalance of tongue image samples in the prior art.

[0008] To achieve the above objectives, the present invention employs the following technical solution: A tongue diagnosis method for rheumatoid arthritis based on information fusion includes the following steps: The original tongue image was converted to the HSV color space, and the optimal segmentation threshold was calculated on the S channel using the Otsu adaptive thresholding algorithm. The tongue image region was extracted by combining morphological opening and closing operations. The tongue region is segmented using an attention U-Net network that incorporates hollow spatial pyramid pooling. The network introduces an attention gating mechanism into the skip connections of the traditional U-Net network and introduces a hollow spatial pyramid pooling module into the deepest layer of feature extraction at the end of the encoder. The tongue coating region is identified from the tongue image using a trained tongue segmentation model. The tongue segmentation model is trained using a transfer learning strategy, and feature selection and sample balancing are performed on the training set data. An adaptive Stacking strategy based on prediction entropy is used to establish a tongue diagnosis recognition model for rheumatoid arthritis. In the tongue diagnosis recognition model for rheumatoid arthritis, the prediction entropy of the base learner is used as the weight of the base learner, and the output probability of the base learner generates a weighted feature input meta-learner to output the prediction result.

[0009] Preferably, the sample balancing process employs a density-weighted boundary mixing sampling strategy, specifically including the following steps: The training set samples are divided into a minority class set and a majority class set; for each minority class sample, its... The nearest neighbors; if the number of samples belonging to the majority class is... satisfy Then the sample is assigned to the boundary set. middle; Calculate the local density factor and its composite weight for each boundary sample in the boundary set, generate the number of composite samples based on the composite weight, and generate composite samples by linear interpolation between the boundary samples and randomly selected minority class neighbor samples; Fuzzy C-means clustering is used to divide the majority class set into C different clusters. Redundant samples in the majority class are removed based on the distance of the sample to the cluster center and a dynamic retention threshold.

[0010] Preferably, the local density factor The calculation formula is: , , in, For boundary samples, It is a positive constant. for With The average Euclidean distance of the nearest minority class neighbor, For the number of minority class neighbor samples, Indicates the first The nearest minority class neighbor sample.

[0011] The composite weight The calculation formula is: , in, These are boundary samples in the boundary set.

[0012] Preferably, the number of synthetic samples The calculation formula is: , in, This is the rounding function. The target is the total number of synthetic samples that need to be generated; The synthetic sample as follows: , in, Let be a random variable, taking values ​​in the range [0.1]. These are randomly selected minority class neighbors.

[0013] Preferably, the majority class set is divided into C distinct clusters using fuzzy C-means clustering, as shown in the following formula: , in, Let C be the objective function for fuzzy C-means clustering. Indicates the first The geometric centers of the clusters, express Belongs to clustering The membership degree, where m is the fuzzy index, takes a value greater than 1. For a majority class set, The number of clusters divided, For the set of cluster centers, This is the membership matrix.

[0014] Preferably, the step of removing redundant samples from the majority class based on the distance from the sample to the cluster center and a dynamic retention threshold specifically includes: Obtain the Euclidean distance from each boundary sample to its cluster center, and establish a dynamic retention threshold for each cluster. If a majority class sample falls within this dense core region, then the condition is met. If so, the boundary sample will be removed.

[0015] Preferably, the method for determining the dynamic retention threshold is as follows: calculate the average Euclidean distance from all majority class samples within the current cluster to their cluster centers, and multiply it by a dynamic adjustment coefficient. The specific formula is as follows: , in, To dynamically retain the threshold, This indicates a clear division to the first The majority class sample set of each cluster, For set The total number of samples in the sample, For the first The geometric centers of the clusters, This is the dynamic adjustment coefficient, with a value range of (0,1).

[0016] Preferably, the adaptive stacking strategy based on predicted entropy specifically includes: Based on the Each base classifier calculates its prediction entropy for predicting the tongue image sample to be tested. The formula is: , in, Indicates the first A base classifier predicts the category to which the tongue image sample belongs. The probability, It is the total number of target categories; Adaptive weights are calculated based on the predicted entropy using a smooth exponential function mapping. The formula is: , in, It is an exponential function with the natural constant e as its base. Represents the total number of base classifiers; Adaptive weights As a feature scaling factor, it is multiplied into the probability distribution features output by the base classifier to generate a weighted feature vector, which is then concatenated and input into the meta-learner for final decision-making.

[0017] The advantages of this invention are as follows: It complements the expressive deficiencies of a single color space through the fusion of features from multiple color spaces, eliminating interference from lighting and device differences. Sample balancing processes balance the data distribution at the feature level, effectively mitigating overfitting. The entropy-weighted stacking strategy dynamically evaluates the reliability of the base learner, significantly improving the accuracy of assisted recognition. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0019] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0020] Figure 2 This is a schematic diagram of a deep feature extraction network structure.

[0021] Figure 3 This is a schematic diagram showing the segmentation and annotation of the tongue. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] like Figure 1 As shown, a tongue diagnosis method for rheumatoid arthritis based on information fusion includes the following steps: S1: Convert the original tongue image to HSV color space, use the Otsu adaptive thresholding algorithm to calculate the optimal segmentation threshold on the S channel, and combine morphological opening and closing operations to extract the tongue image region. S2: The tongue region is segmented using an attention U-Net network that integrates hollow spatial pyramid pooling. The network introduces an attention gating mechanism in the skip connections of the traditional U-Net network and introduces a hollow spatial pyramid pooling module in the deepest layer of feature extraction at the end of the encoder. S3: Use the trained tongue segmentation model to identify the tongue coating region from the tongue image; the tongue segmentation model is trained using a transfer learning strategy, and feature selection and sample balancing are performed on the training set data. S4: An adaptive Stacking strategy based on prediction entropy is used to establish a tongue diagnosis recognition model for rheumatoid arthritis. In the tongue diagnosis recognition model for rheumatoid arthritis, the prediction entropy of the base learner is used as the weight of the base learner. The output probability of the base learner generates a weighted feature input meta-learner to output the prediction result.

[0024] As a refinement of the above embodiment, step S1 converts the original tongue image from the RGB color space to the HSV color space, and uses the Otsu adaptive thresholding algorithm to calculate the optimal segmentation threshold on the S channel (saturation) to initially separate the tongue from the background. Subsequently, morphological opening operations are performed on the binarized mask to remove isolated noise, and closing operations are performed to fill the holes inside the tongue, thereby extracting a complete, realistic, and effective tongue image region.

[0025] As a refinement of the above embodiment, step S2 employs an Attention U-Net network that incorporates dilated spatial pyramid pooling (ASPP-Attention U-Net). This network introduces an attention gating mechanism into the skip connections of the traditional U-Net, which can automatically suppress feature responses in the background region while enhancing the feature weights of the tongue edge region. Furthermore, an ASPP (Dilated Spatial Pyramid Pooling) module is introduced at the end of the encoder to capture multi-scale contextual information, thereby achieving high-precision automatic segmentation of the tongue region.

[0026] As a refinement of the above embodiment, in step S3, before constructing the tongue segmentation model, in order to improve the accuracy of tongue image feature extraction and identification, tongue image preprocessing technology is required to extract the effective information of the tongue image and eliminate errors, thereby improving the efficiency and accuracy of tongue image feature analysis and identification.

[0027] The training process of the tongue segmentation model is as follows: feature extraction and classification are performed on the preprocessed tongue images, and appropriate statistical analysis methods or artificial intelligence algorithms are selected to achieve symptom-assisted identification and risk prediction.

[0028] The deep feature extraction model for tongue diagnosis in rheumatoid arthritis utilizes transfer learning to train a deep feature extraction network that extracts high-order features of the tongue image. Furthermore, it improves accuracy by fusing results from multiple information sources to eliminate errors from individual sources. To enhance the model's ability to capture the features of small lesions in rheumatoid arthritis, the deep feature extraction network employs a deep residual network architecture (SE-ResNet) incorporating a channel attention mechanism. Its complete hierarchical structure includes: 1) Initial convolutional layer: contains a 7x7 convolutional layer with a stride of 2, followed by a batch normalization (BN) layer and a ReLU activation function, used to extract the initial shallow structural features of the tongue image; 2) Initial pooling layer: contains a 3x3 max pooling layer with a stride of 2, used to downsample the initial feature map and expand the receptive field; 3) First Residual Stage: Contains cascaded bottleneck residual blocks. Each residual block contains sequentially 1x1, 3x3, and 1x1 convolutions, combined with batch normalization and ReLU activation functions. A compression and activation (SE) channel attention module is embedded at the end of each residual block, adaptively recalibrating the channel weights based on global average pooling and a two-layer multilayer perceptron. 4) Second residual stage: includes subsequent cascaded bottleneck residual blocks with SE modules, where the first residual block uses a convolution with a stride of 2 for spatial downsampling and increases the channel dimension. 5) Third residual stage: includes bottleneck residual blocks with SE modules that are further cascaded, continue spatial downsampling and expand channel depth to capture high-order semantic features of complex tongue lesions; 6) Fourth residual stage: contains the last set of cascaded bottleneck residual blocks with SE modules, outputting the deepest multidimensional feature map; 7) Global Average Pooling Layer (GAP): Receives the output of the fourth residual stage, performs spatial global average pooling on each feature channel, compresses the spatial dimension to 1x1, generates a one-dimensional high-order semantic feature vector, and greatly reduces the number of parameters. 8) Fully connected layer: Receives a one-dimensional feature vector and performs the final nonlinear transformation; 9) Classification output layer: The output of the fully connected layer is mapped to a probability distribution using the Softmax activation function, and the output is the auxiliary prediction probability or feature representation at the level of the deep feature extraction network.

[0029] Furthermore, the specific internal hierarchical structure of the bottleneck residual block with the SE module includes, in sequence: a) Dimensionality reduction layer: 1x1 convolutional layer, followed by batch normalization (BN) and ReLU activation function, used to compress the channel dimension of the input feature map and reduce computational complexity; b) Spatial extraction layer: 3x3 convolutional layer, followed by BN and ReLU activation functions, used to extract spatial features from the dimensionality-reduced feature map; c) Dimension-up layer: 1x1 convolutional layer followed by BN (not activated immediately), used to restore or expand the channel dimension to match the desired output dimension; d) Compression operation (Squeeze): Spatial global average pooling layer, which compresses the output of the upscaling layer to a size of 1x1 in spatial dimension, generating a channel descriptor vector; e) Excitation: A two-layer multilayer perceptron (MLP) consisting of a dimension-reduced fully connected layer, a ReLU activation function, a dimension-upgraded fully connected layer, and a Sigmoid activation function, outputs a normalized weight value between 0 and 1 for each feature channel; f) Scale layer: The output feature map of the up-dimensional layer is multiplied channel by channel with the channel weights learned from the activation operation to suppress redundant background features and highlight the lesion features of the tongue. g) Residual Connections and Final Activation: Element-wise addition layers merge the original input of the residual block (through skip connections) with the feature map output of the reweighted layer, and finally output the final feature representation of the module through the ReLU activation function.

[0030] Before decision fusion, feature selection and sample balancing are required. The feature selection process uses a filtering feature selection method to screen the fusion features, removing most redundant features. The proposed density-weighted boundary hybrid sampling strategy (DW-BSFCM) operates in two main stages: first, it identifies and oversamples the minority class samples at the boundary. Let the minority class set be... The majority class set is For each minority class sample The algorithm searches for its entire training set. The nearest neighbor. If in this Of the nearest neighbors, the number of samples belonging to the majority class is And meet the conditions Then the sample Partition into boundary set This condition isolates samples located near the decision boundary while explicitly filtering out safe samples and isolated noise. This is crucial for establishing the boundary set. Next, the local density factor of each boundary sample is calculated. To guide the synthesis process. For a given sample Strictly searching for its in a minority of classes Find the nearest neighbors and calculate the average Euclidean distance between it and these neighbors. as follows: , Subsequently, the local density factor is defined as the reciprocal of this average distance, expressed as: , in, For boundary samples, It is a very small positive constant used to prevent division by zero errors. for With The average Euclidean distance of the nearest minority class neighbor, For the number of minority class neighbor samples, Indicates the first The nearest minority class neighbor sample.

[0031] To ensure that the number of synthesized samples is positively correlated with the local density factor, a synthesis weight is assigned to each boundary sample by normalizing the density factor of each boundary sample relative to the sum of the density factors of all boundary samples; the synthesis weight The calculation formula is: , in, Boundary set Boundary samples in.

[0032] Given the target total number of synthetic samples to be generated , by each boundary sample The specific number of synthetic samples derived The calculation is as follows: , in, This is the rounding function.

[0033] Finally, this A new synthetic sample passes through the boundary sample. With randomly selected minority class neighbors The random linear interpolation between them is generated by the following formula: , in, Let be a random variable, taking values ​​in the range [0.1]. These are randomly selected minority class neighbors.

[0034] Following the oversampling phase, the algorithm employs fuzzy C-means (FCM) clustering to perform targeted undersampling (downsampling), effectively removing redundant majority class samples near the cluster centers, thereby constructing a class-balanced, high-quality training set. The FCM algorithm minimizes the objective function to optimize the majority class set. Divide into C distinct clusters: , in, Let C be the objective function for fuzzy C-means clustering. Table 1 The geometric centers of the clusters, express Belongs to clustering Membership degree (limited by) ), where m is the fuzzy index, and its value is greater than 1. For a majority class set, The number of clusters divided, For the set of cluster centers ( ), It is a fuzzy partitioning matrix (or membership matrix).

[0035] Each majority class sample xi is explicitly assigned to its principal cluster. This assignment is determined by the maximum membership degree, that is: , To perform undersampling and mitigate intra-class imbalance, the algorithm evaluates each boundary sample to its respective cluster center. The Euclidean distance. A dynamic retention threshold, i.e., the internal redundancy radius, is established for each cluster. If a majority class sample falls within this dense core region, the condition is met. If a sample is found to be deeply embedded, it is removed. By selectively pruning these deeply embedded samples, this method preserves the structural boundaries of most classes while effectively balancing the overall data distribution at the sample feature level.

[0036] As a refinement of the above embodiments, step S4 adopts an adaptive Stacking strategy to overcome the limitations of the standard voting mechanism.

[0037] General Stacking Architecture: Traditional stacking is a two-stage process. First, multiple differentially differentiated base learners are trained on the dataset. Second, the predictions (probabilities or labels) output by these base learners are concatenated as new features, and this new feature set is used to train a "meta-learner" to make the final prediction. This allows the system to combine the different algorithmic advantages of various models.

[0038] Adaptive improvement based on prediction entropy: Unlike traditional Stacking which uses fixed feature weights, the method proposed in this invention innovatively introduces the "Predictive Entropy" of the base learners as a dynamic measure of classification uncertainty during the decision fusion stage, thereby evaluating the reliability of each base learner in real time. This is particularly important for the diagnosis of rheumatoid arthritis. Due to individual differences, patients' tongue appearance often exhibits "atypical features" or "blurred boundaries" (e.g., tongue coating areas are affected by saliva or reflections).

[0039] Specifically, the The prediction entropy of each base classifier (such as SVM, Random Forest, KNN, or XGBoost in the first layer of base learners) Defined as: , in, Indicates the first A base classifier predicts the category to which the tongue image sample belongs. The probability, It is the total number of target categories (in this embodiment, =2, that is, the rheumatoid arthritis patient class and the healthy normal class).

[0040] When the base classifier struggles to recognize complex lesion features and becomes "confused," the probabilities of each category in its output will tend to be equal (e.g., a 50 / 50 ratio). At this point, the prediction entropy... A large value will be obtained, indicating that the uncertainty is extremely high.

[0041] Subsequently, based on the calculated prediction entropy, adaptive weights are calculated using a smoothing exponential function (similar to Softmax) mapping. : , in, This represents the total number of base classifiers. The weights assigned to base classifiers with high prediction entropy (high uncertainty) are determined through this exponentially decaying mapping. It will decay exponentially.

[0042] This weighting coefficient This feature vector acts as a scaling factor, multiplying the probability distribution features output by the base classifiers to generate a new weighted feature vector. Finally, the weighted features generated by each base classifier are concatenated and input into a second-layer meta-learner (e.g., logistic regression) for the final disease diagnosis decision. Through this mechanism, the strong learner can dynamically and adaptively reduce the contribution of disturbed or ambiguous base models and amplify the decision weights of high-confidence models, based on the diagnostic difficulty of the tongue image being tested. This completely connects the data path between front-end feature extraction and back-end classification decision, significantly improving the robustness and overall diagnostic accuracy of the entire disease-assisted identification system.

[0043] This disclosure also provides a tongue diagnosis recognition device for rheumatoid arthritis based on information fusion, including a processor and a memory. Optionally, the device may further include a communication interface and a bus. The processor, communication interface, and memory can communicate with each other via the bus. The communication interface can be used for information transmission. The processor can call logical instructions in the memory to execute the tongue diagnosis recognition method for rheumatoid arthritis based on information fusion described in the above embodiments.

[0044] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.

[0045] Memory, as a computer-readable storage medium, can be used to store software programs and computer-executable programs, such as the program instructions / modules corresponding to the methods in the embodiments of this disclosure. The processor executes the program instructions / modules stored in the memory to perform functional applications and data processing, thereby realizing the information fusion-based tongue diagnosis recognition method for rheumatoid arthritis described in the above embodiments.

[0046] The memory may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory may include high-speed random access memory and may also include non-volatile memory.

[0047] This disclosure provides a computer-readable storage medium storing computer-executable instructions configured to execute the above-described information fusion-based tongue diagnosis recognition method for rheumatoid arthritis.

[0048] The aforementioned computer-readable storage medium may be a transient computer-readable storage medium or a non-transitory computer-readable storage medium.

[0049] The technical solutions of this disclosure can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes one or more instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in this disclosure. The aforementioned storage medium can be a non-transitory storage medium, including: a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, and other media capable of storing program code. It can also be a transient storage medium.

[0050] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A tongue diagnosis method for rheumatoid arthritis based on information fusion, characterized in that, Includes the following steps: The original tongue image was converted to the HSV color space, and the optimal segmentation threshold was calculated on the S channel using the Otsu adaptive thresholding algorithm. The tongue image region was extracted by combining morphological opening and closing operations. The tongue region is segmented using an attention U-Net network that incorporates hollow spatial pyramid pooling. The network introduces an attention gating mechanism into the skip connections of the traditional U-Net network and introduces a hollow spatial pyramid pooling module into the deepest layer of feature extraction at the end of the encoder. The tongue coating region is identified from the tongue image using a trained tongue segmentation model. The tongue segmentation model is trained using a transfer learning strategy, and feature selection and sample balancing are performed on the training set data. An adaptive Stacking strategy based on prediction entropy is used to establish a tongue diagnosis recognition model for rheumatoid arthritis. In the tongue diagnosis recognition model for rheumatoid arthritis, the prediction entropy of the base learner is used as the weight of the base learner. The output probability of the base learner generates a weighted feature input meta-learner to output the prediction result. The adaptive stacking strategy based on predicted entropy specifically includes: Based on the Each base classifier calculates its prediction entropy for predicting the tongue image sample to be tested. The formula is: ， in, Indicates the first A base classifier predicts the category to which the tongue image sample belongs. The probability, It is the total number of target categories; Adaptive weights are calculated based on the predicted entropy using a smooth exponential function mapping. The formula is: ， in, It is an exponential function with the natural constant e as its base. Represents the total number of base classifiers; Adaptive weights As a feature scaling factor, it is multiplied into the probability distribution features output by the base classifier to generate a weighted feature vector, which is then concatenated and input into the meta-learner for final decision-making.

2. The method for tongue diagnosis in rheumatoid arthritis based on information fusion according to claim 1, characterized in that, The sample balancing process employs a density-weighted boundary mixing sampling strategy, specifically including the following steps: The training set samples are divided into a minority class set and a majority class set; for each minority class sample, its... The nearest neighbors; if the number of samples belonging to the majority class is... satisfy Then the sample is assigned to the boundary set. middle; Calculate the local density factor and its composite weight for each boundary sample in the boundary set, generate the number of composite samples based on the composite weight, and generate composite samples by linear interpolation between the boundary samples and randomly selected minority class neighbor samples; Fuzzy C-means clustering is used to divide the majority class set into C different clusters. Redundant samples in the majority class are removed based on the distance of the sample to the cluster center and a dynamic retention threshold.

3. The method for tongue diagnosis in rheumatoid arthritis based on information fusion according to claim 2, characterized in that, The local density factor The calculation formula is: ， ， in, For boundary samples, It is a positive constant. for With The average Euclidean distance of the nearest minority class neighbor, For the number of minority class neighbor samples, Indicates the first The nearest minority class neighbor sample; The composite weight The calculation formula is: ， in, These are boundary samples in the boundary set.

4. The method for tongue diagnosis in rheumatoid arthritis based on information fusion according to claim 3, characterized in that, The number of synthetic samples The calculation formula is: ， in, This is the rounding function. The target is the total number of synthetic samples that need to be generated; The synthetic sample as follows: ， in, Let be a random variable, taking values ​​in the range [0.1]. These are randomly selected minority class neighbors.

5. The method for tongue diagnosis in rheumatoid arthritis based on information fusion according to claim 2, characterized in that, The method employs fuzzy C-means clustering to divide the majority class set into C distinct clusters, as shown in the following formula: ， in, Let C be the objective function for fuzzy C-means clustering. Indicates the first The geometric centers of the clusters, express Belongs to clustering The membership degree, where m is the fuzzy index, takes a value greater than 1. For a majority class set, The number of clusters divided, For the set of cluster centers, This is the membership matrix.

6. The method for tongue diagnosis in rheumatoid arthritis based on information fusion according to claim 5, characterized in that, The step of removing redundant samples from the majority class based on the distance from the sample to the cluster center and a dynamic retention threshold specifically includes: Obtain the Euclidean distance from each boundary sample to its cluster center, and establish a dynamic retention threshold for each cluster. If a majority class sample falls within a dense core region, then the condition is met. If so, the boundary sample will be removed.

7. The method for tongue diagnosis in rheumatoid arthritis based on information fusion according to claim 6, characterized in that, The method for determining the dynamic retention threshold is as follows: calculate the average Euclidean distance from all majority class samples within the current cluster to their cluster centers, and multiply it by a dynamic adjustment coefficient. The specific formula is as follows: ， in, To dynamically retain the threshold, This indicates a clear division to the first The majority class sample set of each cluster, For set The total number of samples in the sample, For the first The geometric centers of the clusters, This is the dynamic adjustment coefficient, with a value range of (0,1).

8. A tongue diagnosis recognition device for rheumatoid arthritis based on information fusion, comprising a processor and a memory storing program instructions, characterized in that, The processor is configured to execute, when running the program instructions, the information fusion-based tongue diagnosis method for rheumatoid arthritis as described in any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the information fusion-based tongue diagnosis method for rheumatoid arthritis as described in any one of claims 1-7.

Images