Chronic skin ulcer traditional Chinese medicine syndrome type classification system based on pathological feature decoupling

Abstract

The invention belongs to the technical field of medical image processing, and particularly relates to a skin ulcer traditional Chinese medicine syndrome classification system based on deep learning and feature decoupling. Through the displayed feature decoupling and adversarial training mechanism, the deep neural network is forced to simulate the focus in the traditional Chinese medicine syndrome differentiation process, and accurate classification of the chronic skin disease syndrome types is achieved; comprising a shared feature encoder, a color feature decoupling head, a texture feature decoupling head, a color auxiliary discriminator, a texture auxiliary discriminator, a trunk feature extractor and a comprehensive classifier. According to the method, the defect of black box of the existing deep learning model is overcome, and an interpretable diagnosis basis conforming to the traditional Chinese medicine syndrome differentiation logic is provided. Test results show that the method can still keep high accuracy and strong generalization ability under the condition of few training samples, and the application potential of the method in the field of artificial intelligence assisted traditional Chinese medicine diagnosis is shown.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical image processing, and particularly relates to a skin ulcer TCM syndrome type classification system based on deep learning and feature decoupling. BACKGROUND

[0002] Chronic skin ulcer refers to a wound injury that lasts more than four weeks and has no obvious healing trend or a wound that frequently recurs [1], and its diagnosis and treatment are still a big challenge in clinical practice. Chronic skin ulcer seriously affects the quality of life of patients, requires long-term specialized care, and causes huge social and economic costs. It can lead to loss of most skin integrity and even morbidity and mortality. Therefore, it is of great significance to develop intervention treatment strategies for CSU patients.

[0003] For many years, traditional Chinese medicine has treated chronic skin ulcers and accumulated a lot of experience in clinical practice by combining systemic regulation and local treatment, as well as dynamic and combined TCM methods for different syndromes of chronic skin ulcers. "Differential treatment" of TCM skin ulcers is the core content of TCM surgery. The syndrome type (such as blood heat stagnation, qi stagnation and blood stasis, qi and blood deficiency, etc.) directly affects the choice of treatment plan and efficacy.

[0004] Traditional diagnosis relies on physicians to make a comprehensive judgment by observing the color, shape, surface exudates, and other conditions of the ulcer wound. This process is highly dependent on clinical experience and has significant subjectivity and uncertainty. Different TCM physicians have different consistency in judging the same patient's syndrome type, especially in primary medical institutions, inexperienced and newly employed young physicians are prone to syndrome deviation, leading to misdiagnosis or poor efficacy. In addition, traditional diagnosis lacks objective and quantitative standards. Key features such as the degree of redness and swelling around the ulcer, and the consistency of exudates are mostly subjective terms. This experience-driven model not only limits the standardization of TCM skin ulcer diagnosis and treatment, but also hinders the advancement of TCM modernization.

[0005] In recent years, artificial intelligence-assisted diagnosis has become an integral part of modern medicine. It makes objective decisions and evaluations through pre-established algorithms, avoids misjudgments caused by differences in doctors' subjective experience, and significantly improves diagnostic efficiency, becoming a key direction to break through the bottleneck of traditional diagnosis.

[0006] However, there are still significant technical limitations in the application of existing deep learning-based methods in the field of TCM skin ulcers. First, most current mainstream deep learning classification models belong to the "black box" structure. The feature learning process depends on the optimization of the final classification loss. Although certain progress has been made in classification accuracy with improvements in model structure, the invisible and uncontrollable nature of the internal feature extraction and discrimination mechanism of the model makes it lack sufficient explainability. Second, most existing methods focus on the detection and classification of Western medical pathological features, and the data training set is mainly based on Western medical diagnostic criteria. This makes the entire training process generally ignore the integration of TCM feature information such as wound color and surface secretions, which are key information required for TCM syndrome differentiation, resulting in incomplete syndrome differentiation dimensions, difficulty in adapting to abstract syndrome characteristics such as TCM cold, heat, deficiency and excess, and lack of deep analysis of the nature of the syndrome.

[0007] Due to the above reasons, existing deep learning models are difficult to form a corresponding relationship with TCM syndrome differentiation theory, difficult to gain the trust of clinicians, and thus limit their practical application and promotion in the context of TCM skin disease diagnosis and treatment. Therefore, there is an urgent need for a deep learning method that balances high classification accuracy and explainability, allowing the model to fully utilize the TCM feature information in the image and learn the TCM differentiation logic to provide a reliable auxiliary basis for TCM clinical diagnosis.

[0008] The TCM skin ulcer intelligent diagnosis device proposed in this patent achieves a systematic breakthrough in traditional diagnosis mode through multi-dimensional technical innovation. Compared with traditional classification models, this patent utilizes low-level features in ulcer images, which contain TCM feature information such as wound color, shape, surface pus / exudate, etc. These features are converted into quantifiable feature vectors to improve the adaptability and robustness of the model in the field of TCM syndrome differentiation of chronic skin ulcers. SUMMARY

[0009] The purpose of the present invention is to propose a chronic skin ulcer TCM syndrome classification system and method based on pathological feature decoupling with high accuracy and strong generalization ability to improve the adaptability and robustness in the field of TCM syndrome differentiation of chronic skin ulcers.

[0010] Through the explicit feature decoupling and adversarial training mechanism, the deep neural network is forced to mimic the focus points in the TCM syndrome differentiation process (such as color, granulation), achieving accurate classification of chronic skin disease syndromes to guide diagnosis and treatment programs.

[0011] The chronic skin ulcer classification system based on pathological feature decoupling proposed in this invention includes a shared feature encoder, a color feature decoupling head, a texture feature decoupling head, a color auxiliary discriminator, a texture auxiliary discriminator, a main feature extractor, and a comprehensive classifier. Among them:

[0012] The shared feature encoder is configured to extract multi-level shared features from the input image.

[0013] The color feature decoupling head is configured to separate color-related features from the shared features.

[0014] The texture feature decoupling head is configured to separate texture-related features from the shared features.

[0015] The color auxiliary discriminator is configured to perform supervised and adversarial training on the color features.

[0016] The texture auxiliary discriminator is configured to perform supervised and adversarial training on the texture features.

[0017] The backbone feature extractor is configured to further extract deep semantic features from the shared feature map.

[0018] The comprehensive classifier is configured to fuse multi-element features and perform syndrome classification.

[0019] Further,

[0020] (1) The shared feature encoder comprises a starting layer and three residual blocks.

[0021] The starting layer comprises a convolutional layer, a batch normalization layer, a ReLU activation layer, and a max-pooling layer.

[0022] Among the three residual blocks, the first to third residual blocks respectively comprise 2, 3, and 2 residual layers.

[0023] The residual layers adopt a bottleneck structure, which sequentially includes a 1x1 convolution, a 3x3 convolution, a 1x1 convolution, and a residual connection.

[0024] The output feature maps of the first to third residual blocks have channel numbers of 256, 512, and 1024, respectively.

[0025] Here, the shared features refer to features that are used by several parts, specifically, the entire system includes a color decoupling branch, a texture decoupling branch, and a backbone feature extractor, and the inputs of these three parts are directly or indirectly associated with the feature maps output by each layer of the "shared feature encoder"; therefore, these features are essentially shared.

[0026] (2) The color feature decoupling head and the texture feature decoupling head each comprise:

[0027] three convolutional blocks, each comprising a convolutional layer, a batch normalization layer, a ReLU activation function, and a max-pooling layer;

[0028] The output channel numbers are 1024, 512, and 512, respectively.

[0029] (3) The color auxiliary discriminator and the texture auxiliary discriminator each include three residual modules and average pooling layers with stepwise down-sampling; and the final output is a predicted value corresponding to an attribute dimension.

[0030] (4) The backbone feature extractor includes three residual layers and further extracts deep features; the residual structure is the same as that of (1); the input is a feature map obtained by the third residual block of the shared feature encoder, and the channel dimension is 1024; and the output is a feature map with a channel dimension of 2048.

[0031] (5) The comprehensive classifier includes three layers of 3x3 convolution, a global pooling layer and a fully connected layer; the input is a feature map obtained by splicing the outputs of the color feature decoupling head, the texture feature decoupling head and the backbone feature extractor along the channel; and the output is a final classification result.

[0032] A chronic skin ulcer classification method based on the chronic skin ulcer classification system includes the following steps:

[0033] S1: An affected area image of a chronic skin ulcer is obtained, and the size is unified to 3x256x256; and a corresponding TCM syndrome type is labeled, including blood heat stagnation syndrome, qi stagnation and blood stasis syndrome and qi and blood deficiency syndrome;

[0034] S2: Auxiliary feature labels including color attribute vectors and texture attribute vectors are extracted from the original image, which are used as true values for training the decoupling head and the auxiliary discriminator corresponding to the color and the texture, respectively;

[0035] S3: The affected area image is input into the shared feature encoder to obtain a shared feature map;

[0036] S4: The shared feature map is input into the color feature decoupling head and the texture feature decoupling head, respectively, to generate color feature encoding and texture feature encoding;

[0037] S5: The two auxiliary discriminators are used to perform auxiliary supervision and adversarial training on the corresponding feature encoding obtained in step S4, so that each feature encoding only carries information of a specific modality, and feature decoupling is realized;

[0038] S6: The backbone feature extractor is used to extract deep semantic features, also known as backbone features, from the shared feature map;

[0039] S7: The backbone features, color features and texture features are fused;

[0040] S8: The fused features are input into the classifier to output a TCM syndrome type classification result of the chronic skin ulcer.

[0041] Further,

[0042] In step S2, the auxiliary feature label is acquired, wherein the color attribute vector true value is acquired:

[0043] Firstly, the RGB image is converted to HSV color space, where H, S, V represent the hue, saturation, and value of the pixel point, respectively. Among them, the hue represents the color type; the saturation represents the degree of color close to the spectrum color, and the higher the saturation, the deeper and brighter the color; the value represents the brightness of the color in the whole image, which is measured by the maximum pixel value.

[0044] Further, the normalized H-S two-dimensional histogram is calculated; since the brightness only contains the single information of the maximum pixel value, the method focuses on the hue and saturation:

[0045] H is divided into H bin =18 equal small intervals according to the numerical value. Similarly, S is divided into S bin =16 equal small intervals. The 18x16 two-dimensional histogram matrix M is obtained by statistics. The element M(i,j) in the ith row and jth column represents the proportion of the number of pixel points whose H value falls in the ith interval and S value falls in the jth interval to the total number of pixels in the whole image.

[0046] Further, the following color attribute extraction main hue interval index I main , main saturation interval index J main , average hue H avg , hue standard deviation H std , average saturation S avg , saturation standard deviation S std , color distribution entropy E h , color energy W are extracted from the matrix M, a total of 8 color attributes, which are:

[0047] I main =argmax i M(i,j);

[0048]

[0049] E h =-∑ i,j p(i,j)·log2(p(i,j));

[0050]

[0051] For the main hue interval index, The function obtains the row index where the maximum value of the two-dimensional histogram matrix is located. The index represents the color interval where the main color in the image is located.

[0052] For the main saturation interval index, The function gets the column index of the maximum value in the two-dimensional histogram matrix. This index represents the saturation interval where the most dominant color in the image is located.

[0053] For the average hue, H and W represent the height and width of the image, respectively. H(i,j) is the hue value of the pixel at position (i,j) in the original image. This value describes the overall hue tendency of the image.

[0054] For the hue standard deviation, this value describes the degree of fluctuation of the image hue.

[0055] For the average saturation, this value describes the overall saturation tendency of the image.

[0056] For the saturation standard deviation, this value describes the degree of fluctuation of the image saturation.

[0057] For the hue distribution entropy, where p(i,j) is the distribution probability of each hue interval, and is the normalized H(i,j) value, i.e., the distribution probability of each hue interval. The hue distribution entropy measures the degree of chaos of color distribution. The larger the value, the more dispersed the color distribution.

[0058] For the hue energy, this value measures the concentration of color distribution. The larger the value, the more concentrated the color distribution.

[0059] The above color attributes are combined into a color attribute vector A color , which is used as the true value in the training process.

[0060] In step S2, the auxiliary feature label is obtained, where the texture attribute vector true value is obtained:

[0061] Using the Harr wavelet basis function, the image is decomposed into four wavelet subbands, including one low-frequency subband and three high-frequency subbands:

[0062] Low-frequency approximation subband (cA): contains the low-frequency components of the image, representing the overall information of the image;

[0063] Horizontal detail subband (cH): contains high-frequency details in the horizontal direction, mainly representing vertical edge features;

[0064] Vertical detail subband (cV): contains high-frequency details in the vertical direction, mainly representing horizontal edge features;

[0065] Diagonal detail subband (cD): contains high-frequency details in the diagonal direction, mainly representing corner points and particles.

[0066] The energy and entropy values of each subband are calculated, a total of 6 texture attributes, with the specific calculation formula being:

[0067]

[0068] E cH = -∑ k p cH (k) log2(p cH (k));

[0069] E cV = -∑ k p cV (k) log2(p cV (k));

[0070] E cD = -∑ k p cD (k) log2(p cD (k));

[0071] where H c , W c represent the height and width of the sub-band. The height and width of each sub-band are the same, and the value is 1 / 2 of the original image; cH(i,j), cV(i,j), cD(i,j) represent the values of the corresponding sub-band i-th row j-th column, respectively.

[0072] When calculating the entropy of each sub-band, the values of different positions in each sub-band are counted, and are divided into 20 intervals according to the size.

[0073] p cH (k), p cV (k), p cD (k) represent the proportion of each sub-band in the k-th interval, respectively.

[0074] The above texture attributes are combined into a texture attribute vector A texture , which is used as the true value to participate in the training process. The energy and distribution entropy of the wavelet band quantitatively represent the intensity and complexity of the wound texture.

[0075] In step S3, the shared feature encoder includes a starting layer and three residual blocks. The starting layer includes a convolution layer, a batch normalization layer, a ReLU activation layer and a maximum pooling layer, and the first to third residual blocks include 2, 3 and 2 residual layers, respectively. The residual layer adopts a bottleneck structure, which includes 1×1 convolution, 3×3 convolution, 1×1 convolution and residual connection in sequence. The encoder is used for multi-level feature extraction of the image. The input image size is 3×256×256. The sizes of the feature maps f0-f4 output by the starting layer and the first to third residual blocks are 64×64×64, 256×64×64, 512×32×32 and 1024×16×16, respectively.

[0076] In step S4, the color feature decoupling head and the texture feature decoupling head are each composed of three convolutional blocks, each of which includes a 3x3 convolutional layer, a batch normalization layer, a ReLU activation function, and a max pooling layer. The features f2 and f3 are first sampled in the spatial dimension to 64x64 and spliced with f0 and f1. The spliced features pass through the corresponding decoupling head, and finally output the decoupled color feature map f color and the texture feature map f texture .

[0077] In step S5, the color auxiliary discriminator and the texture discriminator include three pre-activation residual blocks, which gradually perform spatial down-sampling through average pooling layers, and interpret the color attribute vector and the texture attribute vector from the color feature map and the texture feature map, respectively:

[0078]

[0079] wherein, and represent the corresponding attribute vectors predicted by the color discriminator and the texture discriminator from the corresponding attribute feature map, respectively, and D color and D texture represent the color discriminator and the texture discriminator, respectively.

[0080] In step S5, the adversarial training phase is represented as follows:

[0081]

[0082] wherein, and represent the corresponding attribute vectors predicted by the color discriminator and the texture discriminator from the non-corresponding attribute feature map, respectively, and GRL(·) represents gradient reversal.

[0083] The loss function of the entire training process includes:

[0084] First, the main task classification loss adopts a cross-entropy function:

[0085]

[0086] wherein y i represents the true label, represents the probability distribution predicted by the classification model, and n represents the number of classification categories

[0087] Second, the auxiliary discrimination loss:

[0088]

[0089] wherein MSE represents the mean square error loss.

[0090] Third, the adversarial loss:

[0091]

[0092] Due to the existence of the gradient reversal layer, the L adv is actually maximized during the training process.

[0093] The total loss is the weighted sum of each loss:

[0094] L total = λ cls L cls + λ attr L attr + λ adv L adv

[0095] where λ cls , λ attr , λ adv represent the weight of each sub-loss. In this scheme, λ cls = 1, λ attr = 0.5, and λ adv = 0.1.

[0096] The present application overcomes the defects of the existing deep learning model black box, and provides an interpretable diagnostic basis that fits the TCM syndrome differentiation logic. Test results show that the present application can maintain high accuracy and strong generalization ability even with fewer training samples, demonstrating its application potential in the field of artificial intelligence assisted TCM diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0097] Figure 1 The classification network structure diagram proposed by the present application.

[0098] Figure 2 The step diagram for extracting color and texture attribute vectors of the present application. DETAILED DESCRIPTION

[0099] The present application will be further described below in conjunction with the accompanying drawings.

[0100] The present application explicitly decouples color and texture features from the original chronic skin ulcer image, forces the deep neural network to mimic the focus points in the TCM syndrome differentiation process (such as wound color, granulation texture condition), and realizes accurate classification of TCM syndromes of chronic skin ulcer (including blood heat stagnation syndrome, qi stagnation and blood stasis syndrome, and deficiency of qi and blood syndrome) to guide the diagnosis and treatment scheme. The specific implementation steps are as follows:

[0101] 1. Image data set preparation

[0102] Images of chronic skin ulcers were acquired, with a uniform size of 3x256x256. The corresponding TCM syndrome types were labeled by professionals, including blood heat stagnation, qi stagnation and blood stasis, and deficiency of qi and blood.

[0103] 2. Auxiliary feature label extraction

[0104] First, color attribute vector extraction: convert the RGB image to the HSV color space, where H, S, and V represent hue, saturation, and brightness, respectively. This method focuses on hue and saturation. Divide the hue H into 18 equal intervals and the saturation S into 16 equal intervals, and generate an 18x16 two-dimensional histogram matrix, with the matrix elements representing the pixel proportion of the corresponding hue and saturation intervals. Extract the following 8 color attributes from the matrix: main hue interval index, main saturation interval index, average hue, hue standard deviation, average saturation, saturation standard deviation, hue distribution entropy, and hue energy. These attributes together form the color attribute vector, which serves as the training true value.

[0105] Second, texture attribute vector extraction

[0106] Perform wavelet transform on the image using the Harr wavelet basis function, decomposing it into four subbands: low-frequency approximation subband (representing the overall information of the image), horizontal detail subband (representing vertical edge features), vertical detail subband (representing horizontal edge features), and diagonal detail subband (representing corner and particle features). Calculate the energy and entropy of each high-frequency subband (horizontal, vertical, and diagonal), resulting in a total of 6 texture attributes.

[0107] 3. Shared feature encoder extracts shallow feature representation

[0108] Input the image into the shared feature encoder. The encoder includes a starting layer and three residual blocks. The starting layer consists of a convolution layer, a batch normalization layer, a ReLU activation function, and a max pooling layer. The three residual blocks each contain 2, 3, and 2 residual layers, respectively, with each residual layer adopting a bottleneck structure consisting of a 1x1 convolution, a 3x3 convolution, a 1x1 convolution, and a residual connection. The encoder outputs four feature maps f0~f3 of different scales, with dimensions of 64x64x64, 256x64x64, 512x32x32, and 1024x16x16, respectively.

[0109] 4. Color and texture feature decoupling

[0110] The feature maps output by the shared feature encoder are upsampled to 64x64 spatial size on feature maps f2 and f3, and are concatenated with feature maps f0 and f1 in the channel dimension. The concatenated features are input into the color feature decoupling head and the texture feature decoupling head, respectively. Each decoupling head consists of three convolutional blocks, each of which includes a 3x3 convolutional layer, a batch normalization layer, a ReLU activation function, and a max pooling layer. The decoupling head finally outputs a 512x8x8 color feature map and a 512x8x8 texture feature map.

[0111] 5. Forward supervised training and adversarial training of color and texture auxiliary discriminators

[0112] In forward supervision, the color auxiliary discriminator predicts the color attribute vector from the color feature map, and the texture auxiliary discriminator predicts the texture attribute vector from the texture feature map. In the adversarial training phase, the texture feature map is input into the color auxiliary discriminator and the color feature map is input into the texture auxiliary discriminator through the gradient reversal layer, forcing the discriminators to predict the attribute vector from the non-corresponding feature map. The training objectives include: first, each auxiliary discriminator accurately predicts the attribute vector from the corresponding feature map (auxiliary discriminator loss). Second, through adversarial training, the features generated by the feature decoupling head are difficult to be predicted by other discriminators, achieving feature decoupling (adversarial loss).

[0113] 6. Feature fusion and classification

[0114] The main features (from the shared encoder), decoupled color features, texture features, and attribute features predicted by the auxiliary discriminators are fused. The fused features are input into the comprehensive classifier, and the traditional Chinese medicine syndrome classification results are output.

[0115] Test results:

[0116] We use the proposed neural network to classify wound images. The wound images are collected by East Hospital Affiliated to Fudan University, and the data set is labeled by three traditional Chinese medicine experts in the field of dermatology. Finally, 423 images are obtained. Among them, there are 218, 72, and 133 images of blood heat stasis syndrome, qi stagnation and blood stasis syndrome, and qi and blood deficiency syndrome, respectively. The training set and test set are divided in the ratio of 8:2 to train and test the above model. To eliminate the randomness of data set division, the model is independently trained for 10 times. In the test phase, the accuracy of the test model for each syndrome category is defined as the proportion of the number of correctly classified images in the total number of images in that category. The average accuracy of the above 10 models is summarized as follows:

[0117] Table 1

[0118] pattern blood-heat stagnation pattern qi stagnation and blood stasis pattern qi and blood deficiency pattern accuracy rate 67.27％ 60.67％ 62.59％

[0119] The test results show that the deep learning model based on pathological feature decoupling proposed in the application has higher accuracy and better generalization ability in the TCM syndrome type classification task of chronic skin ulcers, and shows the adaptability and robustness in the field of TCM syndrome differentiation.

[0120] REFERENCES

[0121] [1]. M. A. Fonder, G. S. Lazarus, D. A. Cowan, B. Aronson-Cook, A. R. Kohli, and A. J. Mamelak, “Treating the chronic wound: a practical approach to the care of nonhealing wounds and wound care dressings,” Journal of the American Academy of Dermatology, vol. 58, no. 2, pp. 185-206, 2008.

Claims

1. A classification system for chronic skin ulcers based on pathological feature decoupling, characterized in that, By employing explicit feature decoupling and adversarial training mechanisms, the deep neural network is forced to mimic the focus points in the TCM diagnostic process, achieving accurate classification of chronic skin disease syndromes. This includes a shared feature encoder, a color feature decoupling head, a texture feature decoupling head, a color-assisted discriminator, a texture-assisted discriminator, a backbone feature extractor, and a comprehensive classifier. Among these components: The shared feature encoder is used to extract multi-level shared features from the input image; The color feature decoupling head is used to separate color-related features from shared features; The texture feature decoupling head is used to separate texture-related features from shared features; The color-assisted discriminator is used for supervised and adversarial training of color features; The texture-assisted discriminator is used for supervised and adversarial training of texture features; The backbone feature extractor is used to further extract deep semantic features from the shared feature map; The comprehensive classifier is used to integrate multiple features and perform syndrome classification.

2. The chronic skin ulcer classification system according to claim 1, characterized in that, (1) The shared feature encoder comprises a start layer and three residual blocks; The initial layer includes a convolutional layer, a batch normalization layer, a ReLU activation layer, and a max pooling layer; Of the three residual blocks, the first to the third residual blocks contain 2, 3, and 2 residual layers, respectively; The residual layer adopts a bottleneck structure, which includes 1×1 convolution, 3×3 convolution, 1×1 convolution and residual connection in sequence; The output feature map channels for the first to third residual blocks are 256, 512, and 1024, respectively. (2) Both the color feature decoupling head and the texture feature decoupling head include: Three convolutional blocks, each containing a convolutional layer, a batch normalization layer, a ReLU activation function, and a max pooling layer; The number of output channels is 1024, 512, and 512 respectively; (3) The color-assisted discriminator and the texture-assisted discriminator both contain three residual modules and an average pooling layer with progressive downsampling; and finally output the predicted value of the corresponding attribute dimension. (4) The backbone feature extractor includes three residual layers to further extract deep features; the residual structure is the same as (1); the input is the feature map obtained by the third residual block of the shared feature encoder, with a channel dimension of 1024; the output is a feature map with a channel dimension of 2048. (5) The comprehensive classifier includes three layers of 3×3 convolution, global pooling layer and fully connected layer; input the feature map after the outputs of the color feature decoupling head, texture feature decoupling head and backbone feature extractor are spliced ​​along the channel; output the final classification result.

3. The chronic skin ulcer classification system according to claim 2, characterized in that, The specific steps for classifying chronic skin ulcers using this system are as follows: S1: Obtain images of the affected area of ​​chronic skin ulcers, with a uniform size of 3×256×256; and label the corresponding TCM syndrome types, including three types: blood heat stagnation syndrome, qi stagnation and blood stasis syndrome, and qi and blood deficiency syndrome; S2: Extract auxiliary feature labels from the original image, including color attribute vector and texture attribute vector, and use them as ground values ​​to train the decoupling head and auxiliary discriminator corresponding to the color and texture respectively; S3: Pass the image of the affected area through the shared feature encoder to obtain a shared feature map; S4: Input the shared feature map into the color feature decoupling head and the texture feature decoupling head respectively to generate color feature code and texture feature code; S5: Use two auxiliary discriminators to perform auxiliary supervision and adversarial training on the corresponding feature codes obtained in step S4, so that each feature code only carries information of a specific modality, thereby achieving feature decoupling; S6: Extract deep semantic features, also known as backbone features, from the shared feature map using the backbone feature extractor; S7: Integrates backbone features, color features, and texture features; S8: Input the fused features into the classifier and output the TCM syndrome classification results of chronic skin ulcers.

4. The chronic skin ulcer classification system according to claim 3, characterized in that, In step S2, the acquisition of the auxiliary feature labels includes the acquisition of the true value of the color attribute vector: First, convert the RGB image to the HSV color space, where H, S, and V represent the hue, saturation, and value of a pixel, respectively. Hue represents the color type; saturation represents how close a color is to the spectral color, with higher saturation resulting in a deeper and more vibrant color; and value represents the brightness of the color in the entire image, measured by the maximum pixel value. Calculate the normalized HS 2D histogram for hue and saturation: H is divided into H based on its numerical value. bin = 18 equal subintervals; similarly, divide S into S0 subintervals. bin = 16 equal intervals; statistically obtain an 18×16 two-dimensional histogram matrix M; where the element M(i,j) in the i-th row and j-th column represents the proportion of pixels whose H value falls in the i-th interval and whose S value falls in the j-th interval to the total number of pixels in the whole image.

5. The chronic skin ulcer classification system according to claim 4, characterized in that, Extract the following color attributes from matrix M: Extract the primary color hue index I. main Main saturation interval index J main Average hue H avg Hue Standard Deviation H std Average saturation S avg Saturation standard deviation S std Color distribution entropy E h Color energy W has a total of 8 color attributes, specifically: I main =argmax i M(i,j); E h =-∑ i,j p(i,j)·log2(p(i,j)); For the primary interval index This function retrieves the row index of the maximum value in a two-dimensional histogram matrix; this index represents the chromaticity range of the most dominant color in the image. For the main saturation interval index This function retrieves the column index of the maximum value in a two-dimensional histogram matrix; this index represents the saturation range of the most dominant color in the image. For the average tone, H and W represent the height and width of the image, respectively; H(i,j) is the tone value of the pixel at position (i,j) in the original image; this value describes the overall tone tendency of the image. For hue standard deviation, this value describes the degree of fluctuation in the image's hue. For average saturation, this value describes the overall saturation tendency of the image; For saturation standard deviation, this value describes the degree of fluctuation in image saturation; For hue distribution entropy, p(i,j) is the probability distribution of each hue interval, and H(i,j) is the normalized value, i.e., the probability distribution of each hue interval; hue distribution entropy measures the degree of disorder in color distribution; the larger the value, the more dispersed the color distribution. For hue energy, this value measures the concentration of color distribution; the larger the value, the more concentrated the color distribution. The above color attributes are combined into a color attribute vector A. color It is used as a truth value in the training process.

6. The chronic skin ulcer classification system according to claim 4, characterized in that, The acquisition of auxiliary feature labels in step S2 includes the acquisition of the true value of the texture attribute vector: Using the Haar wavelet basis function, the image is decomposed into four wavelet subbands, including one low-frequency subband and three high-frequency subbands: Low-frequency approximate subband (cA): Contains the low-frequency components of the image, representing the overall information of the image; Horizontal detail subband (cH): Contains high-frequency details in the horizontal direction, mainly representing vertical edge features; Vertical detail subband (cV): Contains high-frequency details in the vertical direction, mainly representing horizontal edge features; Diagonal detail subband (cD): Contains high-frequency details along the diagonal direction, mainly characterizing corner points and grains; Calculate the energy and entropy values ​​of each subband, a total of 6 texture attributes. The specific calculation formula is as follows: E cH =-∑ k p cH (k)·log2(p cH (k)); E cV =-∑ k p cV (k)·log2(p cV (k)); E cD =-∑ k p cD (k)·log2(p cD (k)); Among them, H c W c Represents the height and width of the sub-band; each sub-band has the same height and width, which is half of the original image; cH(i,j), cV(i,j), and cD(i,j) represent the values ​​in the i-th row and j-th column of the corresponding sub-band, respectively; When calculating the entropy of each sub-band, the values ​​at different positions of each sub-band are statistically analyzed and divided into 20 intervals based on their magnitude; p cH (k), p cV (k), p cD (k) represents the proportion of each sub-band located in the k-th interval; The above texture attributes are combined into a texture attribute vector A. texture It is used as a truth value in the training process.

7. The chronic skin ulcer classification system according to claim 3, characterized in that, In step S3, the shared feature encoder is used to extract multi-level features from the image; the input image size is 3×256×256; the feature maps f0 to f4 output by the first to third residual blocks in the initial layer have the following sizes: 64×64×64, 256×64×64, 512×32×32, and 1024×16×16, respectively.

8. The chronic skin ulcer classification system according to claim 3, characterized in that, In step S4, features f2 and f3 are first sampled to 64×64 in spatial dimension and concatenated with f0 and f1. The concatenated features are then processed by the corresponding decoupling head, and finally output as a 512×8×8 decoupled color feature map f. color and texture feature map f texture .

9. The chronic skin ulcer classification system according to claim 3, characterized in that, In step S5, the color-assisted discriminator and the texture discriminator interpret the color attribute vector and texture attribute vector from the color feature map and the texture feature map, respectively: in, and D represents the corresponding attribute vectors predicted by the color discriminator and the texture discriminator from the corresponding attribute feature maps, respectively. color and D texture These represent the color discriminator and the texture discriminator, respectively.

10. The chronic skin ulcer classification system according to claim 3, characterized in that, The adversarial training phase described in step S5 is represented as follows: in, and represents the corresponding attribute vectors predicted by the color discriminator and the texture discriminator from the non-corresponding attribute feature map, respectively, and GRL(·) represents gradient reversal.

11. The chronic skin ulcer classification system according to claim 9, characterized in that, The loss function during training includes: First, the main task classification loss uses the cross-entropy function: Among them, y i This indicates that it is a genuine label. This represents the probability distribution predicted by the classification model, where n represents the number of classification categories. Second, assist in loss assessment: Where MSE represents the mean squared error loss; Third, combating losses: Because of the gradient inversion layer, L is actually used during training. adv maximize; The total loss is the weighted sum of all losses: L total =λ cls L cls +λ attr L attr +λ adv L adv , Where, λ cls , λ attr , λ adv This represents the weight of each sub-loss.

Images