A method for assessing the degree of recovery of burned skin based on skin image recognition

Abstract

This invention discloses a method for assessing the recovery degree of burned skin based on skin image recognition. The method includes: acquiring visible light and infrared thermal images of a patient with a skin burn, preprocessing them, and fusing them into a skin burn image; performing feature extraction, recognition, and segmentation to obtain an image of the burn area; constructing a static feature set based on burn feature parameters, and constructing a three-dimensional model of the burn wound based on the burn area image and depth image to construct a dynamic feature set; inputting the extracted static and dynamic features into an optimized deep learning-support vector machine fusion model for training, performing final classification of the recovery stage, and calculating the time required for complete healing. The method provided by this invention can not only effectively capture the static feature information of the burn area, but also dynamically monitor the recovery trend of the wound in real time and accurately predict the time required for complete healing, significantly improving the accuracy and personalized assistance of burn diagnosis and treatment.

Description

[Technical Field]

[0001] This invention relates to the field of image recognition technology, and in particular to a method for assessing the degree of recovery of burned skin based on skin image recognition. [Background Technology]

[0002] Burns are a common and severe type of trauma in clinical practice. Their treatment is lengthy and complex, significantly impacting patients' quality of life. Clinically, assessing burn recovery is a crucial basis for treatment decisions.

[0003] Currently, most widely used assessment methods rely on the visual observation and experience of clinicians. This method is significantly subjective, and the assessment results are easily influenced by the assessor's subjective experience, observation skills, and assessment standards, making it difficult to achieve uniformity and standardization. Furthermore, existing technologies lack continuous, quantitative, and dynamic monitoring methods for the burn wound healing process, making it difficult to accurately capture subtle changes in healing and affecting the accurate assessment of burn treatment effectiveness and the implementation of personalized treatment strategies.

[0004] Therefore, it is urgent to develop an objective, accurate, and efficient method for assessing burn recovery. [Summary of the Invention]

[0005] In view of this, embodiments of the present invention provide a method for assessing the degree of recovery of burned skin based on skin image recognition.

[0006] This invention provides a method for assessing the degree of recovery of burned skin based on skin image recognition, the method comprising:

[0007] S1. Acquire visible light and infrared thermal images of patients with skin burns, and fuse them into a skin burn image after preprocessing.

[0008] S2. Use convolutional neural networks and image segmentation networks to extract features, identify and segment skin burn images to obtain images of the burn area;

[0009] S3. Analyze the images of the burned areas to extract burn feature parameters, and construct a static feature set based on the burn feature parameters. Then, classify and grade the burned areas according to the feature parameters.

[0010] S4. Construct a three-dimensional model of the burn wound based on the burn area image and depth image, continuously analyze the burn wound at different times, construct the burn wound recovery curve and calculate the dynamic trend index, and construct a dynamic feature set.

[0011] S5. The extracted static and dynamic features are input into the optimized deep learning-support vector machine fusion model for training, to perform the final classification of the recovery phase, and to calculate the time required for complete healing.

[0012] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein the preprocessing in S1 specifically includes:

[0013] S11. Use high-resolution visible light imaging equipment and infrared thermal imaging equipment to acquire visible light images and infrared thermal images of the burned areas of the skin burn patients, respectively.

[0014] S12. Preprocess the visible light image: use histogram equalization to standardize the image color, use nonlocal mean denoising to reduce noise, and use Laplacian filtering to enhance image details.

[0015] S13. Preprocessing of infrared thermal imaging images: Standardize the temperature information of the image, remove thermal noise by median filtering, and calculate the temperature difference of the burn area using the temperature gradient method.

[0016] As described above and in any possible implementation, a further implementation is provided in which the fusion of the skin burn image in S1 specifically includes:

[0017] S14. Perform Harris corner detection on the preprocessed visible light image and infrared image respectively. The Harris corner response function is defined as: C(x,y)=det(M)-k[trace(M)] 2 , Where C(x,y) represents the corner response value, k represents the empirical constant, M represents the local autocorrelation matrix, det(M) = λ1λ2, trace(M) = λ1 + λ2, λ1 and λ2 are the M eigenvalues ​​of , w(u,v) represents the Gaussian weighted window, I x and I y These represent the gradients of the image in the x and y directions, respectively.

[0018] S15. Extract the feature vectors of corner points from the visible light image and the infrared image respectively, and perform matching using Euclidean distance. Among them, v RGB and v IR represent the feature vectors extracted from the visible light image and the infrared image respectively at the same physical point, where i represents the feature vector dimension index and D represents the Euclidean distance;

[0019] S16. Establish an affine transformation model and use the least squares method to solve for the optimal affine transformation parameters to minimize the registration error, thereby achieving accurate registration of visible light and infrared images. The affine transformation model is defined as follows: (x,y) represents the coordinates of the feature points in the original image, (x′,y′) represents the corresponding coordinates in the target image, a, b, d and e represent the linear transformation parameters, c and f represent the translation parameters, T represents the optimal affine matrix, and E represents the registration error energy function.

[0020] S17. Define temperature anomaly weight W based on local temperature deviation and gradient magnitude. temp (x,y) defines the color anomaly weight W based on the Euclidean distance between the pixel RGB values ​​and the mean of the healthy skin reference color. color (x,y), temperature and color anomaly weights are proportionally fused to obtain the final pixel-level fusion weights to highlight the salience of the burn area, where W temp (x,y)=α|I I ′ R (x,y)-T h |+(1-α)G IR (x,y),

[0021] I' IR (x,y) represents the normalized infrared temperature at pixel (x,y), T h This indicates the reference temperature for healthy skin, G. IR (x,y) represents the magnitude of the temperature gradient, α represents the equilibrium coefficient and α∈[0,1], I R,G,B (x,y) represents the grayscale values ​​of the RGB three channels after detail enhancement. and W represents the average value of the three channels in a healthy skin area. color Indicates the weight of color anomalies. This represents the weighting adjustment coefficient;

[0022] S18. At the same pixel, infrared and RGB grayscale information are fused according to weights to obtain a skin burn image I that combines temperature and texture features. fused (x,y), F(x,y)=W burn (x,y)[I′ IR (x,y)]+[1-W burn (x,y)][I″ gray [x,y)] F(x,y) represents the intermediate values ​​before fusion. This represents the minimum value of the entire graph. I′ represents the maximum value of the entire graph. IR Represents the normalized infrared temperature value, I″ gray This represents the intensity of visible light after grayscale conversion.

[0023] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein S2 specifically includes:

[0024] S21. Feature extraction and recognition of skin burn images are performed using a pre-trained multi-scale convolutional neural network. Specifically, a feature extraction module containing multiple convolutional scales is constructed to extract features from the fused burn image I. fused Feature extraction is performed on (x,y) to obtain feature maps F at different scales. i (x,y), Among them, F i (x, y) represents the convolutional feature map at the i-th scale, where i = 1, 2, 3, corresponding to kernel sizes of 3×3, 5×5, and 7×7, respectively. fused (x,y) represents the fused skin burn image, W i (m,n) represents the weight coefficient at position (m,n) in the i-th scale convolution kernel, b i Let σ(·) represent the bias term of the i-th scale convolution kernel, and let σ(·) represent the nonlinear activation function.

[0025] S22. Construct a fused feature map F by connecting feature maps of different scales through the channel dimension. multi (x,y), and then a 1×1 convolution kernel is used for dimensionality reduction and fusion to obtain the final fused multi-scale feature map F. fusion (x,y), where F multi (x,y)=Concat(F1(x,y),F2(x,y),F3(x,y)),F fusion (x,y)=σ(W f ·F multi (x,y)+b f ), W f b represents the weights of the fused convolutional kernels. f This indicates the fusion of convolution kernel bias terms, and Concat(·) indicates the feature channel concatenation operation;

[0026] S23. The fused multi-scale feature map is segmented into a wound region using a pre-trained image segmentation network to extract the burn region image. Specifically, a U-Net network structure is used, including feature downsampling, feature upsampling, and skip connection structures, to output a probability mask map of the wound region, X. l =Pool(σ(W) l *X l-1 +b l )), Y l =UpSample(σ(W) l *Y l+1 +b l )), Y l =Concat(X) l ,Y l ), P mask (x,y)=Sigmoid(Wout *Y1′(x,y)+b out ), where Pool(·) represents max pooling, UpSample(·) represents upsampling, and bilinear interpolation is used. X l and Y l Let Y represent the l-th layer features of the encoder and decoder, respectively. l ′ represents the feature map fused via skip connections, W out and b out This represents the output convolutional layer parameters, Sigmoid(·) represents the activation function, and P... mask (x,y) represents the probability value of a pixel belonging to the wound region, ranging from [0,1].

[0027] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein S3 specifically includes:

[0028] S31. Obtain skin surface temperature information, calculate the temperature difference between the burned area and healthy skin, and thus extract temperature anomaly features, where T diff (x,y)=|I′ IR (x,y)-T h |,I′ IR (x,y) represents the normalized infrared temperature at pixel (x,y), T h T represents the reference temperature for healthy skin. diff (x,y) represents the temperature difference of the burned area at pixel (x,y);

[0029] S32. Identify granulation tissue regions through texture analysis and calculate the proportion characteristics of granulation tissue, wherein... M mask (x,y) represents the granulation tissue region mask generated by the image segmentation network, T f (x,y) represents the texture features, R gran Indicates the proportion of granulation tissue;

[0030] S33. Identify necrotic tissue through texture feature analysis and calculate the proportion of necrotic tissue, wherein, M′ mask (x,y) represents the mask map of the necrotic tissue region generated by the image segmentation network, A tf (x,y) represents the abnormal features extracted through texture features, R nec Indicates the proportion of necrotic tissue;

[0031] S34. Calculate the epithelialization rate, which reflects the recovery status of the skin surface. A el T represents the area of ​​the epithelial layer. waR represents the total area of ​​the wound. repi Indicates the rate of epithelialization;

[0032] S35. Construct a static feature set based on burn characteristic parameters, wherein the static feature set includes temperature anomaly, granulation tissue ratio, necrotic tissue ratio, and epithelialization rate;

[0033] S36. Based on characteristic parameters, the burned area is preliminarily classified and graded into damage degree and recovery stage. The damage degree is graded into superficial second-degree burns, deep second-degree burns, and third-degree burns. The recovery stage is graded into the inflammatory phase, the proliferative phase, and the remodeling phase. The grading formula is defined as: Class d =f(T) diff ,R gran ,R nec ,R repi ), Class h =g(T) diff ,R gran ,R nec ,R repi f(·) and g(·) represent threshold-based classification functions that classify data using damage and healing features. d and Class h These represent the categories of damage level and recovery stage, respectively.

[0034] In addition to the aspects and any possible implementations described above, an implementation is further provided, wherein S4 specifically includes:

[0035] S41. Obtain depth images of the burned areas of the patient with skin burns, perform preprocessing, and then register and align the depth images with the burned area images.

[0036] S42. Process the depth image to generate a 3D point cloud, where P 3D (x,y)=(x,y,I depth (x,y)), I depth (x,y) represents the depth value of this pixel, P 3D (x,y) represents the coordinates of the corresponding pixel in three-dimensional space;

[0037] S43. For burn area images and depth images at different time points, point clouds at different time points are processed by time series processing, and point cloud registration and alignment are performed using the ICP algorithm.

[0038] S44. Use the Marching Cubes algorithm to connect the registered point clouds to generate a 3D mesh model, and then optimize it. M 3D (t)=Mesh({P 3D (x,y)}), M 3D(t) represents the three-dimensional burn wound model at time point t, and Mesh({·}) represents converting the point cloud into a mesh model;

[0039] S45. By analyzing the 3D model at different times, a wound recovery curve is constructed, and the recovery rate is calculated. M mask (x,y,t) represents the masking layer covering the burn wound area at time t, A re (t) represents the area or volume of the burn wound at time point t, V rate Indicates the recovery rate;

[0040] S46. Construct a dynamic feature set based on the wound recovery curve and dynamic trend indicators. The dynamic feature set includes the burn wound recovery area or volume and the recovery rate.

[0041] In addition to the aspects and any possible implementations described above, a further implementation is provided in which the deep learning-support vector machine fusion model optimized by inputting the static and dynamic features extracted in S5 is trained to perform the final classification in the recovery phase, specifically including:

[0042] S51. Construct a deep learning-support vector machine fusion model. Use the concatenation of static and dynamic feature sets as input, employ a multi-scale convolutional neural network to extract high-dimensional fusion features, and then feed them into a support vector machine classifier for classification.

[0043] S52. Define a training set. The samples in the training set are pre-classified and graded to obtain soft labels. The training set is used to train the model. In the early stage of training, the soft labels are incorporated into the loss function in a smoothing penalty manner. The loss function is specifically defined as the weighted sum of hinge loss and soft label loss.

[0044] S53. Through an end-to-end training strategy, the parameters of CNN and SVM models are alternately optimized. The optimization goal is to minimize the classification error. Cross-validation is used to adjust the network structure parameters of CNN, the kernel function parameters of SVM, and the soft label penalty coefficient to obtain the best classification effect.

[0045] S54. After training, the fusion model is used to perform the final classification of the recovery stage and output the classification results of the burn recovery stage.

[0046] In addition to the aspects and any possible implementations described above, a further implementation is provided in which the calculation of the time required for complete healing in S5 specifically includes:

[0047] Define a healing time formula and calculate the time required for complete healing using this formula, where... T healP represents the number of remaining days predicted for complete healing at the current moment. inf T represents the probability that the current burn, as output by the fusion model, belongs to the inflammatory phase. inf P represents the average baseline time required for a wound in the inflammatory phase to fully heal. pro T represents the probability that the current burn output by the fusion model belongs to the proliferative phase. pro P represents the average baseline time required for a wound in the proliferative phase to fully heal. rem T represents the probability that the current burn output by the fusion model belongs to the remodeling phase. rem D represents the average baseline time required for a wound in the remodeling phase to fully heal. se Indicates the moderating factor for the degree of injury, where superficial second-degree burns D se =0.8, deep second-degree burn D se =1.0, third-degree burn D se =1.5, H adj E represents the burn recovery adjustment factor. rate Indicates the current epithelialization rate. This represents the absolute value of the rate of change of the burn wound area. The absolute value of the rate of change of burn wound volume is represented by k, where ΔT represents the temperature difference between the burned area and the healthy area. t These represent dynamic trend indicators, with a, b, c, d, and e representing coefficients.

[0048] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein the method further includes:

[0049] S61. The patient end and / or medical node end acquire images of the burn area through a smart mobile device with a built-in visible light and infrared thermal imaging camera and upload them to the cloud server.

[0050] S62. The cloud server automatically performs burn image recognition to assess the degree of recovery of burned and damaged skin, automatically classifies the burn recovery stage, calculates the healing time, and sends the analysis results to the medical institution.

[0051] S63. The medical institution formulates and updates the burn care plan in real time based on the received analysis results, and feeds it back to the patient and / or medical node through the cloud server. At the same time, the medical institution requests remote expert assistance for diagnosis and treatment as needed.

[0052] S64. The remote expert terminal remotely receives images collected by the patient terminal and / or medical node terminal, nursing plans and burn analysis results from the medical institution terminal, and provides professional guidance opinions, which are fed back to the medical institution terminal, patient terminal and / or medical node terminal through the cloud server.

[0053] S65. Data transmission between the patient terminal and / or medical node terminal, medical institution terminal and remote expert terminal is conducted through interactive communication, data encryption storage and recording via cloud server.

[0054] One of the above technical solutions has the following beneficial effects:

[0055] This invention proposes a method for assessing the recovery degree of burned skin based on skin image recognition. This method integrates multi-source image data from visible light and infrared thermal imaging, employing advanced multi-scale convolutional neural network technology, image segmentation technology, 3D point cloud modeling technology, and deep learning-support vector machine fusion classification technology to objectively and accurately quantify and assess the degree of damage and recovery process of burned skin. The assessment method provided by this invention not only effectively captures the static feature information of the burned area but also constructs a 3D model of the burn wound based on the burned area image and depth image, dynamically monitors the recovery trend of the wound in real time, and accurately predicts the time required for complete healing. This method can significantly improve the accuracy and personalized assistance of burn diagnosis and treatment, providing reliable decision-making basis for patients and medical personnel. [Attached Image Description]

[0056] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 This is a flowchart illustrating S1-S5 of the method for assessing the degree of recovery of burned skin based on skin image recognition provided in this embodiment of the invention.

[0058] Figure 2 This is a flowchart illustrating methods S11-S13 provided in an embodiment of the present invention;

[0059] Figure 3 This is a flowchart illustrating methods S14-S18 provided in an embodiment of the present invention;

[0060] Figure 4 This is a flowchart illustrating methods S21-S23 provided in an embodiment of the present invention;

[0061] Figure 5 This is a flowchart illustrating methods S31-S36 provided in an embodiment of the present invention;

[0062] Figure 6 This is a flowchart illustrating methods S41-S46 provided in an embodiment of the present invention;

[0063] Figure 7 This is a flowchart illustrating methods S51-S54 provided in an embodiment of the present invention;

[0064] Figure 8 This is a flowchart illustrating methods S61-S65 provided in an embodiment of the present invention.

Detailed Implementation Methods

[0065] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0066] Please refer to Figure 1 This is a flowchart illustrating S1-S5 of a method for assessing the degree of recovery of burned skin based on skin image recognition, as provided in an embodiment of the present invention. Figure 1 As shown, the method includes the following steps:

[0067] S1. Acquire visible light and infrared thermal images of patients with skin burns, and fuse them into a skin burn image after preprocessing.

[0068] S2. Use convolutional neural networks and image segmentation networks to extract features, identify and segment skin burn images to obtain images of the burn area;

[0069] S3. Analyze the images of the burned areas to extract burn feature parameters, and construct a static feature set based on the burn feature parameters. Then, classify and grade the burned areas according to the feature parameters.

[0070] S4. Construct a three-dimensional model of the burn wound based on the burn area image and depth image, continuously analyze the burn wound at different times, construct the burn wound recovery curve and calculate the dynamic trend index, and construct a dynamic feature set.

[0071] S5. The extracted static and dynamic features are input into the optimized deep learning-support vector machine fusion model for training, to perform the final classification of the recovery phase, and to calculate the time required for complete healing.

[0072] This invention integrates multi-source image data from visible light and infrared thermal imaging, employing advanced multi-scale convolutional neural network technology, image segmentation technology, 3D point cloud modeling technology, and deep learning-support vector machine fusion classification technology to objectively and accurately quantify and assess the degree of damage and recovery process of burned skin. The assessment method provided by this invention not only effectively captures the static feature information of the burned area but also constructs a 3D model of the burn wound based on the burned area image and depth image, dynamically monitors the wound's recovery trend in real time, and accurately predicts the time required for complete healing. This method significantly improves the accuracy and personalized assistance of burn diagnosis and treatment, providing reliable decision-making basis for patients and medical personnel.

[0073] In a preferred embodiment of the present invention, such as Figure 2 As shown, the preprocessing in S1 specifically includes:

[0074] S11. Use high-resolution visible light imaging equipment and infrared thermal imaging equipment to acquire visible light images and infrared thermal images of the burned areas of the skin burn patients, respectively.

[0075] S12. Preprocess the visible light image: use histogram equalization to standardize the image color, use nonlocal mean denoising to reduce noise, and use Laplacian filtering to enhance image details.

[0076] S13. Preprocessing of infrared thermal imaging images: Standardize the temperature information of the image, remove thermal noise by median filtering, and calculate the temperature difference of the burn area using the temperature gradient method.

[0077] This invention improves image contrast and detail by applying histogram equalization, nonlocal mean denoising, and Laplacian filtering to visible light images, and performs temperature standardization and median filtering on infrared thermal images to remove thermal noise and highlight temperature differences. This significantly improves image quality, reduces the interference of noise on analysis results during image processing, and effectively enhances the accuracy of subsequent feature extraction and image fusion.

[0078] In a preferred embodiment of the present invention, such as Figure 3 As shown, the process of fusing the skin burn image in S1 specifically includes:

[0079] S14. Perform Harris corner detection on the preprocessed visible light image and infrared image respectively. The Harris corner response function is defined as: C(x,y)=det(M)-k[trace(M)] 2 , Where C(x,y) represents the corner response value, k represents the empirical constant, M represents the local autocorrelation matrix, det(M) = λ1λ2, trace(M) = λ1 + λ2, λ1 and λ2 are the M eigenvalues ​​of , w(u,v) represents the Gaussian weighted window, I x and I y These represent the gradients of the image in the x and y directions, respectively.

[0080] S15. Extract the feature vectors of corner points from the visible light image and the infrared image respectively, and perform matching using Euclidean distance. Among them, v RGB and v IR represent the feature vectors extracted from the visible light image and the infrared image respectively at the same physical point, where i represents the feature vector dimension index and D represents the Euclidean distance;

[0081] S16. Establish an affine transformation model and use the least squares method to solve for the optimal affine transformation parameters to minimize the registration error, thereby achieving accurate registration of visible light and infrared images. The affine transformation model is defined as follows: (x,y) represents the coordinates of the feature points in the original image, (x′,y′) represents the corresponding coordinates in the target image, a, b, d and e represent the linear transformation parameters, c and f represent the translation parameters, T represents the optimal affine matrix, and E represents the registration error energy function.

[0082] S17. Define temperature anomaly weight W based on local temperature deviation and gradient magnitude. temp (x,y) defines the color anomaly weight W based on the Euclidean distance between the pixel RGB values ​​and the mean of the healthy skin reference color. color (x,y), temperature and color anomaly weights are proportionally fused to obtain the final pixel-level fusion weights to highlight the salience of the burn area, where W temp (x,y)=α|I′ IR (x,y)-T h |+(1-α)G IR (x,y),

[0083] I' IR (x,y) represents the normalized infrared temperature at pixel (x,y), T h This indicates the reference temperature for healthy skin, G. IR (x,y) represents the magnitude of the temperature gradient, α represents the equilibrium coefficient and α∈[0,1], I R,G,B (x,y) represents the grayscale values ​​of the RGB three channels after detail enhancement. and W represents the average value of the three channels in a healthy skin area. color Indicates the weight of color anomalies. This represents the weighting adjustment coefficient;

[0084] S18. At the same pixel, infrared and RGB grayscale information are fused according to weights to obtain a skin burn image I that combines temperature and texture features. fused (x,y), F(x,y)=W burn (x,y)[I′ IR (x,y)]+[1-W burn (x,y)][I″ gray [x,y)] F(x,y) represents the intermediate values ​​before fusion. This represents the minimum value of the entire graph. I′ represents the maximum value of the entire graph. IR Represents the normalized infrared temperature value, I″ gray This represents the intensity of visible light after grayscale conversion.

[0085] This invention utilizes Harris corner detection and Euclidean distance matching of feature vectors, combined with affine transformation solved by least squares method, to achieve high-precision registration and fusion of visible light and infrared images, resulting in a high-quality burn fusion image that combines temperature and texture features. By employing dual anomaly weights of temperature and color for precise fusion, the fusion image more clearly highlights the burn area, greatly improving the accuracy of burn area identification and the reliability of image analysis.

[0086] In a preferred embodiment of the present invention, such as Figure 4 As shown, S2 specifically includes:

[0087] S21. Feature extraction and recognition of skin burn images are performed using a pre-trained multi-scale convolutional neural network. Specifically, a feature extraction module containing multiple convolutional scales is constructed to extract features from the fused burn image I. fused Feature extraction is performed on (x,y) to obtain feature maps F at different scales. i (x,y), Among them, F i (x, y) represents the convolutional feature map at the i-th scale, where i = 1, 2, 3, corresponding to kernel sizes of 3×3, 5×5, and 7×7, respectively. fused (x,y) represents the fused skin burn image, W i (m,n) represents the weight coefficient at position (m,n) in the i-th scale convolution kernel, b i Let σ(·) represent the bias term of the i-th scale convolution kernel, and let σ(·) represent the nonlinear activation function.

[0088] S22. Construct a fused feature map F by connecting feature maps of different scales through the channel dimension. multi (x,y), and then a 1×1 convolution kernel is used for dimensionality reduction and fusion to obtain the final fused multi-scale feature map F.fusion (x,y), where F multi (x,y)=Concat(F1(x,y),F2(x,y),F3(x,y)),F fusion (x,y)=σ(W f ·F multi (x,y)+b f ), W f b represents the weights of the fused convolutional kernels. f This indicates the fusion of convolution kernel bias terms, and Concat(·) indicates the feature channel concatenation operation;

[0089] S23. The fused multi-scale feature map is segmented into a wound region using a pre-trained image segmentation network to extract the burn region image. Specifically, a U-Net network structure is used, including feature downsampling, feature upsampling, and skip connection structures, to output a probability mask map of the wound region, X. l =Pool(σ(W) l *X l-1 +b l )), Y l =UpSample(σ(W) l *Y l+1 +b l )), Y l =Concat(X) l ,Y l ), P mask (x,y)=Sigmoid(W out *Y1′(x,y)+b out ), where Pool(·) represents max pooling, UpSample(·) represents upsampling, and bilinear interpolation is used. X l and Y l Let Y represent the l-th layer features of the encoder and decoder, respectively. l ′ represents the feature map fused via skip connections, W out and b out This represents the output convolutional layer parameters, Sigmoid(·) represents the activation function, and P... mask (x,y) represents the probability value of a pixel belonging to the wound region, ranging from [0,1].

[0090] This invention employs a pre-trained multi-scale convolutional neural network to extract feature maps at different convolutional kernel scales and performs channel concatenation and convolutional dimensionality reduction, capturing information about the wound at different spatial detail levels, significantly improving the ability to capture complex burn features in images. Simultaneously, the U-Net network structure is used to perform encoding-decoding operations and skip connections on the multi-scale fused feature maps, accurately segmenting the burn region and its internal structure, ensuring the spatial coherence and edge detail integrity of the segmentation results. The segmentation mask provides a high-precision wound region, greatly improving the accuracy of wound region detection and localization, and ensuring the stability and reliability of the evaluation results.

[0091] It should be noted that, depending on the needs, the multi-scale convolutional neural network and image segmentation network can be designed in this application, or other existing networks that can achieve the above functions can be used.

[0092] In a preferred embodiment of the present invention, such as Figure 5 As shown, S3 specifically includes:

[0093] S31. Obtain skin surface temperature information, calculate the temperature difference between the burned area and healthy skin, and thus extract temperature anomaly features, where T diff (x,y)=|I′ IR (x,y)-T h |,I′ IR (x,y) represents the normalized infrared temperature at pixel (x,y), T h T represents the reference temperature for healthy skin. diff (x,y) represents the temperature difference of the burned area at pixel (x,y);

[0094] S32. Identify granulation tissue regions through texture analysis and calculate the proportion characteristics of granulation tissue, wherein... M mask (x,y) represents the granulation tissue region mask generated by the image segmentation network, T f (x,y) represents the texture features, R gran Indicates the proportion of granulation tissue;

[0095] S33. Identify necrotic tissue through texture feature analysis and calculate the proportion of necrotic tissue, wherein, M′ mask (x,y) represents the mask map of the necrotic tissue region generated by the image segmentation network, A tf (x,y) represents the abnormal features extracted through texture features, R nec Indicates the proportion of necrotic tissue;

[0096] S34. Calculate the epithelialization rate, which reflects the recovery status of the skin surface. Ael T represents the area of ​​the epithelial layer. wa R represents the total area of ​​the wound. repi Indicates the rate of epithelialization;

[0097] S35. Construct a static feature set based on burn characteristic parameters, wherein the static feature set includes temperature anomaly, granulation tissue ratio, necrotic tissue ratio, and epithelialization rate;

[0098] S36. Based on characteristic parameters, the burned area is preliminarily classified and graded into damage degree and recovery stage. The damage degree is graded into superficial second-degree burns, deep second-degree burns, and third-degree burns. The recovery stage is graded into the inflammatory phase, the proliferative phase, and the remodeling phase. The grading formula is defined as: Class d =f(T) diff ,R gran ,R nec ,R repi ), Class h =g(T) diff ,R gran ,R nec ,R repi f(·) and g(·) represent threshold-based classification functions that classify data using damage and healing features. d and Class h These represent the categories of damage level and recovery stage, respectively.

[0099] Based on segmentation results and normalized temperature maps, this invention calculates the temperature difference between the burn area and healthy skin, quantifies the degree of temperature abnormality, quantitatively assesses the pixel ratio of granulation tissue and necrotic tissue through texture analysis to reflect the degree of wound tissue reconstruction and necrosis, extracts epithelialization rate parameters to reflect the regeneration of surface cells in the wound, and combines the four parameters of temperature abnormality, granulation tissue ratio, necrotic tissue ratio, and epithelialization rate into a static feature vector to achieve a comprehensive quantitative description of the current state of the wound. Furthermore, based on a threshold formula, it performs preliminary classification and grading of the degree of damage and recovery stage. Through the establishment of a specific and refined set of static features, the classification and grading assessment of burn areas becomes more scientific and standardized, significantly improving the accuracy of assessment and the personalization of treatment strategies.

[0100] In a preferred embodiment of the present invention, such as Figure 6 As shown, S4 specifically includes:

[0101] S41. Obtain depth images of the burned areas of the patient with skin burns, perform preprocessing, and then register and align the depth images with the burned area images.

[0102] S42. Process the depth image to generate a 3D point cloud, where P 3D(x,y)=(x,y,I depth (x,y)), I depth (x,y) represents the depth value of this pixel, P 3D (x,y) represents the coordinates of the corresponding pixel in three-dimensional space;

[0103] S43. For burn area images and depth images at different time points, point clouds at different time points are processed by time series processing, and point cloud registration and alignment are performed using the ICP algorithm.

[0104] S44. Use the Marching Cubes algorithm to connect the registered point clouds to generate a 3D mesh model, and then optimize it. M 3D (t)=Mesh({P 3D (x,y)}), M 3D (t) represents the three-dimensional burn wound model at time point t, and Mesh({·}) represents converting the point cloud into a mesh model;

[0105] S45. By analyzing the 3D model at different times, a wound recovery curve is constructed, and the recovery rate is calculated. M mask (x,y,t) represents the masking layer covering the burn wound area at time t, A re (t) represents the area and / or volume of the burn wound region at time point t, V rate Indicates the recovery rate;

[0106] S46. Construct a dynamic feature set based on the wound recovery curve and dynamic trend indicators. The dynamic feature set includes the rate of change of burn wound recovery area and / or volume, as well as the recovery rate.

[0107] This invention preprocesses point clouds from multi-time-phase depth images and registers them using the ICP algorithm to generate a continuous 3D mesh model. The Marching Cubes algorithm is then used to construct and optimize 3D wound models at each time point, accurately calculating the wound volume or area at each moment. A wound recovery curve is constructed, and dynamic indicators such as recovery rate and volume change trends are calculated, forming a dynamic feature vector. This vector intuitively reflects the temporal evolution of wound healing, accurately monitors the dynamic changes and recovery trends of the wound, and effectively improves the ability to finely monitor the healing process of burn wounds, providing more intuitive and accurate dynamic trend indicators. The depth images can be obtained using a structured light depth camera or images acquired from two different perspectives using stereo vision methods.

[0108] In a preferred embodiment of the present invention, such as Figure 7As shown, the deep learning-support vector machine fusion model, optimized by inputting the static and dynamic features extracted in S5, is trained to perform the final classification in the recovery phase, specifically including:

[0109] S51. Construct a deep learning-support vector machine fusion model. Use the concatenation of static and dynamic feature sets as input, employ a multi-scale convolutional neural network to extract high-dimensional fusion features, and then feed them into a support vector machine classifier for classification.

[0110] S52. Define a training set. The samples in the training set are pre-classified and graded to obtain soft labels. The training set is used to train the model. In the early stage of training, the soft labels are incorporated into the loss function in a smoothing penalty manner. The loss function is specifically defined as the weighted sum of hinge loss and soft label loss.

[0111] S53. Through an end-to-end training strategy, the parameters of CNN and SVM models are alternately optimized. The optimization goal is to minimize the classification error. Cross-validation is used to adjust the network structure parameters of CNN, the kernel function parameters of SVM, and the soft label penalty coefficient to obtain the best classification effect.

[0112] S54. After training, the fusion model is used to perform the final classification of the recovery stage and output the classification results of the burn recovery stage.

[0113] This invention concatenates static and dynamic features and inputs them into a multi-scale CNN to extract high-dimensional fused features, which are then fed into an SVM classifier for final classification during the recovery phase, achieving high-precision identification of the inflammation, proliferative, and remodeling phases. An efficient fusion classification model combining deep learning and support vector machines is established, significantly improving the accuracy of the final classification of burn recovery stages. By alternately optimizing the parameters of the CNN and SVM models, the accuracy and reliability of the classification results are achieved, providing a highly reliable recovery assessment. During the model training phase, a soft-label smoothing penalty and an end-to-end alternating optimization strategy are introduced to collaboratively adjust the CNN network structure, SVM kernel function, and penalty coefficients, ensuring optimal performance of the classification model.

[0114] Specifically, the method for training a CNN-SVM classification model based on the fusion of static and dynamic features is as follows:

[0115] Constructing the input feature set:

[0116] The static feature set F constructed above s and dynamic feature set F d The concatenation forms a unified input feature vector X, defined as follows: Where, ΔT diff Indicates an abnormal temperature, R gran R represents the proportion of granulation tissue. nec R represents the proportion of necrotic tissue.repi Indicates the epithelialization rate, Indicates the rate of change of area. V represents the rate of change of volume. rate Indicates the recovery rate;

[0117] Constructing a fusion CNN-SVM model:

[0118] The CNN structure is as follows:

[0119] Input layer: Input normalized feature vectors X∈R d The dimension is d; convolutional layers: two one-dimensional convolutional layers (1D-CNN) are used, with kernel sizes of 3 and 5, to extract multi-scale high-dimensional features H1 = ReLU(Conv1D). k＝3 (X; W1, b1)), H2 = ReLU (Conv1D k＝5 (X;W2,b2)), where W1 represents the kernel parameters of the first convolutional layer, b1 represents the bias parameters of the first convolutional layer, W2 represents the kernel parameters of the second convolutional layer, and b2 represents the bias parameters of the second convolutional layer; Pooling layer: Max pooling further extracts salient features, H pool1 =MaxPool(H1), H pool2 =MaxPool(H2); Concatenation layer: concatenates the pooled features and outputs a fused feature vector Z, Z = [H2]. pool1 H pool2 ]∈R m m represents the dimension; fully connected layer: further compresses the feature dimension Z. feature Z feature =ReLU(W fc Z+b fc ), W fc b represents the weight parameters of the fully connected layer. fc Indicates the bias parameters of the fully connected layer;

[0120] The obtained Z feature As input, it is fed into the SVM classifier;

[0121] Define an SVM classifier that maps high-dimensional features to the classification space using a kernel function:

[0122] The radial basis function (RBF) is used, and the RBF kernel function is defined as K(Z) i Z j )=exp(-γ||Z i -Z j || 2 ), where γ represents the RBF kernel function width hyperparameter, Z i and Z j This represents the feature vector extracted by the CNN;

[0123] The decision function is defined as: α i Let y represent the Lagrange multiplier of the SVM classifier. i represents the true class label of the sample, and b represents the SVM classifier bias term;

[0124] Apply soft label smoothing penalty using initial classification labels:

[0125] Preliminary soft label y is defined based on the preliminary classification and grading results. soft , representing the probability that a sample belongs to each category;

[0126] The loss function is defined as a weighted combination of the SVM hinge loss and the soft-label smoothing penalty loss. Where w represents the SVM weights, C represents the SVM normalization coefficients, and y i Let φ represent the true classification label, φ represent the mapping function, and λ represent the soft label penalty coefficient. This indicates that the CNN extracts features from the i-th training sample;

[0127] The training steps for the fused CNN-SVM model are as follows: randomly initialize CNN parameters, fix SVM parameters, optimize CNN parameters using Adam, fix CNN parameters again, update SVM parameters using the SMO algorithm until the model's loss on the validation set no longer decreases; cross-validate CNN structure parameters, SVM parameters, and smoothing penalty coefficients.

[0128] After training, given a new feature vector, CNN features are extracted, and an SVM classifier is used to predict the recovery phase, namely the inflammation phase, the proliferative phase, and the remodeling phase.

[0129] It should be noted that the CNN-SVM model in this application can be the one designed in this application, or it can be other existing models that can achieve the above functions.

[0130] In a preferred embodiment of the present invention, calculating the time required for complete healing in step S5 specifically includes:

[0131] Define a healing time formula and calculate the time required for complete healing using this formula, where... T heal P represents the number of remaining days predicted for complete healing at the current moment. inf T represents the probability that the current burn, as output by the fusion model, belongs to the inflammatory phase. inf P represents the average baseline time required for a wound in the inflammatory phase to fully heal. pro T represents the probability that the current burn output by the fusion model belongs to the proliferative phase. pro P represents the average baseline time required for a wound in the proliferative phase to fully heal.rem T represents the probability that the current burn output by the fusion model belongs to the remodeling phase. rem D represents the average baseline time required for a wound in the remodeling phase to fully heal. se Indicates the moderating factor for the degree of injury, where superficial second-degree burns D se =0.8, deep second-degree burn D se =1.0, third-degree burn D se =1.5, H adj E represents the burn recovery adjustment factor. rate Indicates the current epithelialization rate. This represents the absolute value of the rate of change of the burn wound area. The absolute value of the rate of change of burn wound volume is represented by k, where ΔT represents the temperature difference between the burned area and the healthy area. t These represent dynamic trend indicators, with a, b, c, d, and e representing coefficients.

[0132] Based on classification probability and preset baseline time, this invention constructs a formula for predicting the complete healing time of burns. It comprehensively considers multi-dimensional information such as damage degree adjustment factors, epithelialization rate, volume change rate and temperature trend, so as to achieve accurate quantitative prediction of complete healing time, which greatly improves the accuracy of complete healing time prediction and ensures accurate formulation of treatment plans and monitoring of efficacy.

[0133] In a preferred embodiment of the present invention, such as Figure 8 As shown, the method further includes:

[0134] S61. The patient end and / or medical node end acquire images of the burn area through a smart mobile device with a built-in visible light and infrared thermal imaging camera and upload them to the cloud server.

[0135] S62. The cloud server automatically performs burn image recognition to assess the degree of recovery of burned and damaged skin, automatically classifies the burn recovery stage, calculates the healing time, and sends the analysis results to the medical institution.

[0136] S63. The medical institution formulates and updates the burn care plan in real time based on the received analysis results, and feeds it back to the patient and / or medical node through the cloud server. At the same time, the medical institution requests remote expert assistance for diagnosis and treatment as needed.

[0137] S64. The remote expert terminal remotely receives images collected by the patient terminal and / or medical node terminal, nursing plans and burn analysis results from the medical institution terminal, and provides professional guidance opinions, which are fed back to the medical institution terminal, patient terminal and / or medical node terminal through the cloud server.

[0138] S65. Data transmission between the patient terminal and / or medical node terminal, medical institution terminal and remote expert terminal is conducted through interactive communication, data encryption storage and recording via cloud server.

[0139] This invention constructs a complete collaborative system involving the patient and / or medical node, medical institution, remote expert, and cloud server ends, enabling real-time transmission and sharing of burn image data and assessment results. This allows medical institutions to respond quickly and dynamically adjust nursing plans. The addition of remote diagnosis and expert guidance functions significantly improves the utilization rate of medical resources and the timeliness and effectiveness of treatment, promotes the development of precision medicine and telemedicine, and greatly improves the patient's treatment experience and outcomes.

[0140] Furthermore, in a preferred embodiment of the present invention, a system for assessing the degree of recovery of burn-damaged skin based on skin image recognition is also proposed. This system includes various modules for implementing the above-mentioned method for assessing the degree of recovery of burn-damaged skin based on skin image recognition.

[0141] It is understandable that the processing module and the overlay module can be any conventional processor, controller, microcontroller, or state machine. The processor can also be implemented by combining components of the analysis system, such as digital signal processors and microprocessors, multiple microprocessors, one or more microprocessors combined with a digital signal processor core, or any other similar configuration.

[0142] Furthermore, in a preferred embodiment of the present invention, a storage medium is provided on which a computer program is stored. When the computer program is invoked and executed, it performs a method for assessing the degree of recovery of burn-damaged skin based on skin image recognition as described above. The storage medium can be a magnetic medium, such as a floppy disk, hard disk, or magnetic tape; an optical medium, such as a DVD; or a semiconductor medium, such as a solid-state drive (SSD).

[0143] The foregoing has shown and described the basic principles, main features, and advantages of this invention. It should be noted that, to maintain dimensional consistency, all parameters involved in the formulas of this invention have undergone dimensionless preprocessing. Those skilled in the art should understand that this invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely the principles of the invention. Various changes and modifications can be made to this invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the invention as claimed. The scope of protection claimed by this invention is defined by the appended claims and their equivalents.

Claims

1. A method for assessing the degree of recovery of burned skin based on skin image recognition, characterized in that, The method includes: S1. Acquire visible light and infrared thermal images of patients with skin burns, and fuse them into a skin burn image after preprocessing. S2. Use convolutional neural networks and image segmentation networks to extract features, identify and segment skin burn images to obtain images of the burn area; S3. Analyze the burn area image to extract burn feature parameters, and construct a static feature set based on the burn feature parameters. The static feature set includes temperature abnormality, granulation tissue ratio, necrotic tissue ratio and epithelialization rate. Perform preliminary classification and grading of the burn area according to the feature parameters. S4. Construct a three-dimensional model of the burn wound based on the burn area image and depth image, continuously analyze the burn wound at different times, construct the burn wound recovery curve and calculate the dynamic trend index, and construct a dynamic feature set. S5. The extracted static and dynamic features are input into the optimized deep learning-support vector machine fusion model for training, to perform the final classification of the recovery phase, and to calculate the time required for complete healing. S4 specifically includes: S41. Obtain depth images of the burned areas of the patient with skin burns, perform preprocessing, and then register and align the depth images with the burned area images. S42. Process the depth image to generate a 3D point cloud, wherein... , This represents the depth value of the pixel. Represents the coordinates of the corresponding pixel in three-dimensional space; S43. For burn area images and depth images at different time points, point clouds at different time points are processed by time series processing, and point cloud registration and alignment are performed using the ICP algorithm. S44. The Marching Cubes algorithm is used to connect the registered point clouds to generate a 3D mesh model, which is then optimized. , This represents a three-dimensional burn wound model at time point t. This indicates that the point cloud is converted into a mesh model; S45. By analyzing the 3D model at different times, a wound recovery curve is constructed, and the recovery rate is calculated. , , This represents the mask covering the burn wound area at time point t. This indicates the area or volume of the burn wound at time point t. Indicates the recovery rate; S46. Construct a dynamic feature set based on the wound recovery curve and dynamic trend indicators, wherein the dynamic feature set includes the burn wound recovery area or volume and the recovery rate. The calculation of the time required for complete healing in S5 specifically includes: Define a healing time formula and calculate the time required for complete healing using this formula, where... , , This indicates the number of days remaining to achieve full healing as predicted at the current moment. This represents the probability that the current burn, as output by the fusion model, belongs to the inflammatory phase. This represents the average baseline time required for a wound in the inflammatory phase to fully heal. This represents the probability that the current burn, as output by the fusion model, belongs to the proliferative phase. This represents the average baseline time required for a wound in the proliferative phase to fully heal. This indicates the probability that the current burn, as output by the fusion model, belongs to the remodeling phase. This represents the average baseline time required for a wound in the remodeling phase to fully heal. Indicating damage severity modulators, among which, superficial second-degree burns... Deep second-degree burns Third-degree burns , Indicates the adjustment factor for burn recovery degree. Indicates the current epithelialization rate. This represents the absolute value of the rate of change of the burn wound area. This represents the absolute value of the rate of change of burn wound volume. This indicates the temperature difference between the burned area and the healthy area. Indicates a dynamic trend indicator. , , , and Represents the coefficient.

2. The method according to claim 1, characterized in that, The preprocessing in S1 specifically includes: S11. Use high-resolution visible light imaging equipment and infrared thermal imaging equipment to acquire visible light images and infrared thermal images of the burned area of ​​the skin burn patient, respectively. S12. Preprocess the visible light image: use histogram equalization to standardize the image color, use nonlocal mean denoising to reduce noise, and use Laplacian filtering to enhance image details. S13. Preprocessing of infrared thermal imaging images: Standardize the temperature information of the image, remove thermal noise by median filtering, and calculate the temperature difference of the burn area using the temperature gradient method.

3. The method according to any one of claims 1 or 2, characterized in that, The fused skin burn image in S1 specifically includes: S14. Perform Harris corner detection on the preprocessed visible light image and infrared image respectively, whereby the Harris corner response function is defined as: , ,in, Indicates the corner response value. Represents an empirical constant. Represents the local autocorrelation matrix. , , and For Eigenvalues This represents a Gaussian weighted window. and These represent the images in Gradient of direction; S15. Extract the feature vectors of corner points from the visible light image and the infrared image respectively, and perform matching using Euclidean distance. ,in, and These represent the feature vectors extracted from the visible light image and the infrared image, respectively, at the same physical point. Indicates the feature vector dimension index. Indicates Euclidean distance; S16. Establish an affine transformation model and use the least squares method to solve for the optimal affine transformation parameters to minimize the registration error, thereby achieving accurate registration of visible light and infrared images. The affine transformation model is defined as follows: , , Represents the coordinates of feature points in the original image. Indicates the coordinates corresponding to the target image. , , and Represents the parameters of the linear transformation. and Indicates the translation parameter. Denotes the optimal affine matrix. Represents the energy function of registration error; S17. Define temperature anomaly weights based on local temperature deviation and gradient magnitude. Color anomaly weights are defined based on the Euclidean distance between pixel RGB values ​​and the mean of a healthy skin reference color. The weights for temperature and color anomalies are proportionally fused to obtain the final pixel-level fusion weights, which highlight the salience of the burn area. , , ， ， Represents the normalized pixel Infrared temperature, Indicates the reference temperature for healthy skin. Indicates the magnitude of the temperature gradient. Represents the balance coefficient and , This indicates the RGB three-channel grayscale after detail enhancement. , and This represents the average value of the three channels in a healthy skin area. Indicates the weight of color anomalies. This represents the weighting adjustment coefficient; S18. At the same pixel, infrared and RGB grayscale information are fused according to weights to obtain a skin burn image that combines temperature and texture features. , , , This represents the intermediate value before merging. This represents the minimum value of the entire graph. This represents the maximum value in the entire graph. This represents the normalized infrared temperature value. This represents the intensity of visible light after grayscale conversion.

4. The method according to claim 1, characterized in that, S2 specifically includes: S21. Feature extraction and recognition of skin burn images are performed using a pre-trained multi-scale convolutional neural network. Specifically, a feature extraction module containing multiple convolutional scales is constructed to process the fused burn image. Feature extraction was performed to obtain feature maps at different scales. , ,in, Indicates the first Each scale of convolutional feature map The corresponding convolution kernel sizes are respectively , This represents a fused image of skin burns. Indicates the first Location in each scale convolution kernel The weighting coefficients, Indicates the first Bias terms of convolution kernels at each scale Represents a nonlinear activation function; S22. Construct a fused feature map by connecting feature maps of different scales through the channel dimension. Then, a 1×1 convolution kernel is used for dimensionality reduction and fusion to obtain the final fused multi-scale feature map. ,in, , , Indicates the weights of the fused convolutional kernels. Indicates the fusion convolution kernel bias term. This indicates a cascaded operation of feature channels; S23. The fused multi-scale feature map is segmented using a pre-trained image segmentation network to extract the burn region image. Specifically, a U-Net network structure is used, including feature downsampling, feature upsampling, and skip connection structures, to output a probability mask map of the burn region. , , , ,in, This indicates a max pooling operation. This indicates an upsampling operation using bilinear interpolation. and This indicates that the encoder and decoder respectively represent the first... Layer features, This represents the feature map fused via skip connections. and This indicates the output convolutional layer parameters. This represents the activation function. This represents the probability value of a pixel belonging to the wound area, within a certain range. .

5. The method according to claim 4, characterized in that, S3 specifically includes: S31. Obtain skin surface temperature information, calculate the temperature difference between the burned area and healthy skin, and extract temperature anomaly features. , Represents the normalized pixel Infrared temperature, This indicates the reference temperature for healthy skin. Indicates the burned area in pixels Temperature difference at the location; S32. Identify granulation tissue regions through texture analysis and calculate the proportion characteristics of granulation tissue, wherein... , This represents a mask image of the granulation tissue region generated by an image segmentation network. Represents texture features, Indicates the proportion of granulation tissue; S33. Identify necrotic tissue through texture feature analysis and calculate the proportion of necrotic tissue, wherein, , This represents a mask image of the necrotic tissue region generated by an image segmentation network. This represents the abnormal features extracted through texture features. Indicates the proportion of necrotic tissue; S34. Calculate the epithelialization rate, which reflects the recovery status of the skin surface. , This represents the area of ​​the epithelial layer. This indicates the total area of ​​the wound. Indicates the rate of epithelialization; S35. Construct a static feature set based on burn characteristic parameters, wherein the static feature set includes temperature anomaly, granulation tissue ratio, necrotic tissue ratio, and epithelialization rate; S36. Based on characteristic parameters, the burn area is preliminarily classified and graded into damage degree and recovery stage. The damage degree is graded into superficial second-degree burns, deep second-degree burns, and third-degree burns. The recovery stage is graded into the inflammatory phase, the proliferative phase, and the remodeling phase. The grading formula is defined as follows: , , and This represents a threshold-based classification function that uses damage and healing features for classification. and These represent the categories of damage level and recovery stage, respectively.

6. The method according to claim 1, characterized in that, The deep learning-support vector machine fusion model, optimized by inputting the static and dynamic features extracted in S5, is trained to perform the final classification in the recovery phase, specifically including: S51. Construct a deep learning-support vector machine fusion model. Use the concatenation of static and dynamic feature sets as input, employ a multi-scale convolutional neural network to extract high-dimensional fusion features, and then feed them into a support vector machine classifier for classification. S52. Define a training set. The samples in the training set are pre-classified and graded to obtain soft labels. The training set is used to train the model. In the early stage of training, the soft labels are incorporated into the loss function in a smoothing penalty manner. The loss function is specifically defined as the weighted sum of hinge loss and soft label loss. S53. Through an end-to-end training strategy, the parameters of CNN and SVM models are alternately optimized. The optimization goal is to minimize the classification error. Cross-validation is used to adjust the network structure parameters of CNN, the kernel function parameters of SVM, and the soft label penalty coefficient to obtain the best classification effect. S54. After training, the fusion model is used to perform the final classification of the recovery stage and output the classification results of the burn recovery stage.

7. The method according to claim 1, characterized in that, The method further includes: S61. The patient end and / or medical node end acquire images of the burn area through a smart mobile device with a built-in visible light and infrared thermal imaging camera and upload them to the cloud server. S62. The cloud server automatically performs burn image recognition to assess the degree of recovery of burned and damaged skin, automatically classifies the burn recovery stage, calculates the healing time, and sends the analysis results to the medical institution. S63. The medical institution formulates and updates the burn care plan in real time based on the received analysis results, and feeds it back to the patient and / or medical node through the cloud server. At the same time, the medical institution requests remote expert assistance for diagnosis and treatment as needed. S64. The remote expert terminal remotely receives images collected by the patient terminal and / or medical node terminal, nursing plans and burn analysis results from the medical institution terminal, and provides professional guidance opinions, which are fed back to the medical institution terminal, patient terminal and / or medical node terminal through the cloud server. S65. Data transmission between the patient terminal and / or medical node terminal, medical institution terminal and remote expert terminal is conducted through interactive communication, data encryption storage and recording via cloud server.

Images