A method of wound healing time prediction

Abstract

The present application belongs to the field of wound assessment, and specifically refers to a wound healing time prediction method. The method comprises information collection, primary classification, information complementation, training stage, secondary classification, statistical wound characteristic data and prediction of healing time. The present scheme introduces temperature characteristic data, quantifies inflammatory cells and inflammatory mediators, obtains dressing information and wound tissue type, so that the neural network model can fully consider the strong correlation of the characteristic data to learn and predict the wound healing time. Through spatial alignment and information fusion of two-dimensional wound images and three-dimensional wound images, the neural network learns the surface features of the wound and the anatomical information of the soft tissue, improves the estimation of the severity of the deep wound and the accuracy of the time prediction of the neural network, and at the same time, uses simulated images to replace paired training images to overcome the dependence on annotated registration data in the traditional supervised learning method.

Description

Technical Field

[0001] This invention relates to the field of wound assessment, specifically to a method for predicting wound healing time. Background Technology

[0002] One method for predicting wound healing time is to use neural networks and image registration technology to assess wound healing time, effectively assisting physicians in quantifying the recovery period of injured patients.

[0003] Among the existing approximate solutions, such as CN115426939A, a machine learning system and method for wound assessment, healing prediction and treatment, this solution uses machine learning technology and light detection elements to achieve the technical effect of non-invasively and non-invasively classifying tissue areas at the wound site. However, it has the problem of using only optical images as analysis data and not taking into account the role of wound dressings, inflammatory cells and inflammatory mediators in wound healing.

[0004] In addition, traditional deep learning algorithms can predict wound healing time. CN114663699B proposes a method for high-precision identification of wound tissue type and prediction of wound healing time. This method uses the XGBoost model. By adding new wound features to the training set and iteratively training the XGBoost model, it can accurately determine the wound tissue type and get rid of the technical effect of relying on doctors' experience. However, it has the technical problem that the input image is only a two-dimensional wound image after the patient's debridement. For wounds with indistinct surface but deep trauma, the prediction error is large.

[0005] Therefore, traditional wound healing time prediction models have technical problems, such as using only optical images as analysis data without considering the role of wound dressings, inflammatory cells, and inflammatory mediators in wound healing; and the input images are only two-dimensional wound images after debridement of patients, which has a large prediction error for wounds with indistinct surfaces but deep trauma. Summary of the Invention

[0006] To address the above issues and overcome the shortcomings of existing technologies, this invention provides a method for predicting wound healing time. Traditional wound healing time prediction models rely solely on optical images for analysis, neglecting the roles of wound dressings, inflammatory cells, and inflammatory mediators in wound healing. This method employs a partial differential equation model to quantify the effects of inflammatory cells and mediators on vascular endothelial growth factor concentration and the density of newly formed capillary tips. It introduces thermograms to obtain wound temperature characteristics, acquires dressing information and wound tissue type, and uses the Kruskal-Wallis test to screen the correlation between these characteristics and the degree of wound healing. This allows the neural network model to comprehensively consider highly correlated characteristics to predict wound healing time. Learning and Prediction: Addressing the technical challenge of large prediction errors for wounds with indistinct surfaces but deep wounds that are difficult to observe directly, where the input image is only a two-dimensional image of the patient's debridement, this solution introduces three-dimensional wound images from MRI scans to display the location, depth, and surrounding tissue damage of deep wounds. By spatially aligning the two-dimensional and three-dimensional wound images, image registration is achieved. This allows the neural network to learn wound surface features while simultaneously training its understanding of soft tissue anatomy, improving the accuracy of the neural network's assessment of the extent of deep wound damage and prediction of healing time. Furthermore, by using simulated images instead of paired training images, the reliance on annotated registration data in traditional supervised learning methods is overcome.

[0007] The technical solution adopted by this invention is as follows: This invention provides a method for predicting wound healing time, which includes the following steps:

[0008] Step S1: Information acquisition, specifically, acquiring two-dimensional wound images through an RGB camera and an infrared thermal imaging camera, and using magnetic resonance imaging technology to obtain a three-dimensional MRI image containing soft tissue anatomical information, and recording the MRI three-dimensional image as a three-dimensional wound image;

[0009] Step S2: Primary classification, specifically, using a deep neural network to perform binary classification on the two-dimensional wound image, dividing the two-dimensional wound image into wound region and non-wound region;

[0010] Step S3: Information complementarity for image registration, ensuring that the structure of the 3D wound image can be accurately located in the 2D wound image. Specifically, deep learning technology is used to construct a deep learning neural network, which spatially aligns and fuses the 2D and 3D wound images, and outputs a fused and registered image.

[0011] Step S4: Training phase;

[0012] Step S5: Secondary classification for identifying tissue types. Specifically, based on the segmentation results of the primary classification, an image segmentation network based on a multilayer perceptron is used to segment the fused and registered image into fixed-size pixel blocks. Each pixel block is classified to identify the tissue type contained in each pixel block; prediction scores for different tissue types contained in each pixel block are calculated.

[0013] Step S6: Collect wound feature data to extract and filter wound feature data;

[0014] Step S7: Predict healing time. Specifically, the selected wound feature data and the predicted score are used as combined features. A Gaussian process regression model is used with the combined features as input data to predict the number of days required for wound healing.

[0015] Furthermore, in step S3, the information complementarity specifically includes the following steps:

[0016] Step S31: Simulation image generation, used to generate simulated 2D images to expand the training dataset. Specifically, an encoder is used to generate a simulated 2D wound image corresponding to the 3D wound image. The simulated 2D wound image corresponding to the 3D wound image is recorded as the simulated 2D image, and the real 2D wound image corresponding to the 3D wound image is recorded as the real 2D image.

[0017] Step S32: Reduce degrees of freedom, specifically by using a preset performance control technique as a constraint to limit the rotation and translation required for the 3D wound image to transform from the initial registration estimate to the optimal registration estimate.

[0018] The initial registration estimate represents the initial position and orientation of the three-dimensional wound image;

[0019] The optimal registration estimate represents the position and orientation when the three-dimensional wound image and the two-dimensional wound image are perfectly aligned, i.e., the final position and orientation;

[0020] Step S33: Feature adaptation, used to reduce the difference between simulated 2D images and real 2D images at the feature level. Specifically, it involves using a generative adversarial network to reduce differences in feature structure and image style. The difference in feature structure specifically refers to the difference in feature content representation between the simulated 2D image and the real 2D image, and the difference in image style specifically refers to the difference in visual appearance between the simulated 2D image and the real 2D image.

[0021] The three-dimensional wound image and the corresponding real 2D image are taken at different times, with different imaging devices used, and with different body positions of the injured person. These real 2D images are denoted as unpaired real images. A simulated two-dimensional wound image is generated based on the unpaired real images. The simulated two-dimensional wound image generated based on the unpaired real images is denoted as an unpaired simulated image.

[0022] Using the unpaired real graph and the unpaired simulated graph, the deep learning neural network performs feature domain adaptation operations through unsupervised learning.

[0023] Furthermore, in step S33, the feature domain adaptive operation specifically includes the following steps:

[0024] Step S331: Reduce feature structure differences to increase the realism of the simulated 2D image generated by the encoder. Specifically, calculate the encoder's adversarial loss function using the following formula: ;

[0025] In the formula, This represents the adversarial loss function of the encoder. and These represent simulated 2D images and real 2D images, respectively. Indicates input data, This represents the expected value calculated by sampling feature maps from a simulated 2D image as input data. This represents the expected value calculated by sampling feature maps from a real 2D image as input data. Representing feature maps, Indicates the discriminator;

[0026] The adversarial loss function is minimized through backpropagation;

[0027] Step S332: Reduce image style differences. The specific steps are as follows:

[0028] Step S3321: Calculate feature correlation to evaluate the linear relationship between different feature dimensions of the simulated 2D image. Calculate the cross-correlation matrix between features of different style versions of the simulated 2D image generated from the same 3D wound image. The calculation formula is as follows: ;

[0029] In the formula, The elements of the cross-correlation matrix are represented. Indicates the number of samples. Indicates the sample index. and This represents the features of two simulated 2D images derived from different stylistic versions of the same 3D wound image. and They represent the first time. In the samples, the The Dimensions and the stated The The value of the dimension;

[0030] Step S3322: Reduce redundancy, which reduces redundancy between features without increasing the differences in the feature structure. Specifically, it calculates the redundancy loss function using the following formula: ;

[0031] In the formula, Represents the redundancy loss function. Represents the cross-correlation matrix, This represents the diagonal elements of the cross-correlation matrix. and Both represent feature dimension indices. Represents the invariant term. Indicates the redundant weighting factor. Indicates the redundancy reduction term;

[0032] By minimizing the redundancy loss function, i.e., minimizing the off-diagonal elements of the cross-correlation matrix, redundancy between features is reduced;

[0033] Step S333: Feature dimensionality reduction; Principal component analysis is used to reduce the dimensionality of high-dimensional features;

[0034] Step S334: Style transfer, used to reduce registration errors caused by style differences. Specifically, through comparative learning, the gap between the simulated 2D image and the real 2D image is narrowed, and style transfer from the real 2D image style to the simulated 2D image style is achieved.

[0035] Furthermore, in step S4, the training phase specifically includes the following steps:

[0036] Step S41: Pre-training, using a simulated 2D image generated from the 3D wound image to pre-train the deep learning neural network, which learns how to register the 3D wound image with the simulated 2D image;

[0037] Step S42: Joint training, using simulated 2D images and real 2D images generated from 3D wound images to jointly train the deep learning neural network, which processes both simulated 2D images and real 2D images simultaneously;

[0038] Step S43: The trained deep learning neural network outputs a fused and registered image.

[0039] Furthermore, in step S6, the statistical wound feature data specifically includes the following steps:

[0040] Step S61: Based on the fused registration image and the result of the first classification, the wound region is segmented from the fused registration image to obtain the temperature distribution of the wound region. Histogram equalization technology is used to enhance the contrast of the wound region. Machine vision technology is used to calculate the wound area. Based on the temperature distribution of the wound region in the fused registration image, temperature feature data is calculated. The temperature feature data includes the average temperature, median, standard deviation, skewness, and kurtosis of the wound region at each time point.

[0041] Step S62: Construct a mathematical model using partial differential equations. The mathematical model is used to quantify the effects of inflammatory cells and inflammatory mediators on the concentration of vascular endothelial growth factor and the density of newly formed capillary tips.

[0042] Step S63: Medication record, specifically, collecting the frequency of wound dressing and the usage of dressing products, and recording it as dressing information;

[0043] Step S64: Record the temperature characteristic data, the mathematical model, and the dressing information as wound characteristic data;

[0044] Step S65: Statistical analysis is used to screen wound feature data. Specifically, the Kruskal-Wallis test is used with a 95% confidence interval to analyze the correlation between the wound feature data and the wound healing status, and wound feature data related to the wound healing status are screened.

[0045] This invention provides a method for predicting wound healing time. The beneficial effects achieved by this invention using the above scheme are as follows:

[0046] (1) In view of the technical problem that traditional wound healing time prediction models only use optical images as analysis data and do not take into account the role of wound dressing, inflammatory cells and inflammatory mediators in wound healing, this scheme adopts a partial differential equation model to quantify the effect of inflammatory cells and inflammatory mediators on vascular endothelial growth factor concentration and new capillary tip density, introduces heat map to obtain wound temperature feature data, obtain dressing information and wound tissue type, and screens the correlation between the above feature data and wound healing degree through Kruskal-Wallis test, so that the neural network model can fully consider the highly correlated feature data to learn and predict wound healing time;

[0047] (2) In response to the technical problem that the input image is only a two-dimensional wound image of the patient after debridement, which is difficult to observe directly for wounds with indistinct surface but deep trauma, the prediction error is large, this solution introduces three-dimensional wound images from MRI scans to show the location, depth and damage of the surrounding tissues of deep wounds. By spatially aligning the two-dimensional wound image with the three-dimensional wound image, image registration is achieved, which enables the neural network to learn the surface features of the wound while also training it to learn and understand the anatomical information of soft tissues, thereby improving the accuracy of the neural network in estimating the severity of deep wounds and predicting the time. At the same time, simulated images are used to replace paired training images, overcoming the dependence on annotated registration data in traditional supervised learning methods. Attached Figure Description

[0048] Figure 1 A flowchart illustrating a method for predicting wound healing time provided by the present invention;

[0049] Figure 2 This is a flowchart illustrating step S3;

[0050] Figure 3 This is a flowchart illustrating step S33;

[0051] Figure 4 This is a flowchart illustrating step S4.

[0052] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0053] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0054] Example 1, see Figures 1 to 4 The present invention provides a method for predicting wound healing time, the method comprising the following steps:

[0055] Step S1: Information acquisition, specifically, acquiring two-dimensional wound images through an RGB camera and an infrared thermal imaging camera, and using magnetic resonance imaging technology to obtain a three-dimensional MRI image containing soft tissue anatomical information, and recording the MRI three-dimensional image as a three-dimensional wound image;

[0056] Step S2: Primary classification, specifically, using a deep neural network to perform binary classification on the two-dimensional wound image, dividing the two-dimensional wound image into wound region and non-wound region;

[0057] Step S3: Information complementarity for image registration, ensuring that the structure of the 3D wound image can be accurately located in the 2D wound image. Specifically, deep learning technology is used to construct a deep learning neural network, which spatially aligns and fuses the 2D and 3D wound images, and outputs a fused and registered image.

[0058] Step S4: Training phase;

[0059] Step S5: Secondary classification for identifying tissue types. Specifically, based on the segmentation results of the primary classification, an image segmentation network based on a multilayer perceptron is used to segment the fused and registered image into fixed-size pixel blocks. Each pixel block is classified to identify the tissue type contained in each pixel block; prediction scores for different tissue types contained in each pixel block are calculated.

[0060] Step S6: Collect wound feature data to extract and filter wound feature data;

[0061] Step S7: Predict healing time. Specifically, the selected wound feature data and the predicted score are used as combined features. A Gaussian process regression model is used with the combined features as input data to predict the number of days required for wound healing.

[0062] Example 2, see Figures 1 to 3 This embodiment is based on the above embodiment. In step S3, the information complementarity specifically includes the following steps:

[0063] Step S31: Simulation image generation, used to generate simulated 2D images to expand the training dataset. Specifically, an encoder is used to generate a simulated 2D wound image corresponding to the 3D wound image. The simulated 2D wound image corresponding to the 3D wound image is recorded as the simulated 2D image, and the real 2D wound image corresponding to the 3D wound image is recorded as the real 2D image.

[0064] Step S32: Reduce degrees of freedom, specifically by using a preset performance control technique as a constraint to limit the rotation and translation required for the 3D wound image to transform from the initial registration estimate to the optimal registration estimate.

[0065] The initial registration estimate represents the initial position and orientation of the three-dimensional wound image;

[0066] The optimal registration estimate represents the position and orientation when the three-dimensional wound image and the two-dimensional wound image are perfectly aligned, i.e., the final position and orientation;

[0067] Step S33: Feature adaptation, used to reduce the difference between simulated 2D images and real 2D images at the feature level. Specifically, it involves using a generative adversarial network to reduce differences in feature structure and image style. The difference in feature structure specifically refers to the difference in feature content representation between the simulated 2D image and the real 2D image, and the difference in image style specifically refers to the difference in visual appearance between the simulated 2D image and the real 2D image.

[0068] The three-dimensional wound image and the corresponding real 2D image are taken at different times, with different imaging devices used, and with different body positions of the injured person. These real 2D images are denoted as unpaired real images. A simulated two-dimensional wound image is generated based on the unpaired real images. The simulated two-dimensional wound image generated based on the unpaired real images is denoted as an unpaired simulated image.

[0069] Using the unpaired real graph and the unpaired simulated graph, the deep learning neural network performs feature domain adaptation operations through unsupervised learning.

[0070] Example 3, see Figures 1 to 3 This embodiment is based on the above embodiment. In step S33, the feature domain adaptive operation specifically includes the following steps:

[0071] Step S331: Reduce feature structure differences to increase the realism of the simulated 2D image generated by the encoder. Specifically, calculate the encoder's adversarial loss function using the following formula: ;

[0072] In the formula, This represents the adversarial loss function of the encoder. and These represent simulated 2D images and real 2D images, respectively. Indicates input data, This represents the expected value calculated by sampling feature maps from a simulated 2D image as input data. This represents the expected value calculated by sampling feature maps from a real 2D image as input data. Representing feature maps, Indicates the discriminator;

[0073] The adversarial loss function is minimized through backpropagation;

[0074] Step S332: Reduce image style differences. The specific steps are as follows:

[0075] Step S3321: Calculate feature correlation to evaluate the linear relationship between different feature dimensions of the simulated 2D image. Calculate the cross-correlation matrix between features of different style versions of the simulated 2D image generated from the same 3D wound image. The calculation formula is as follows: ;

[0076] In the formula, The elements of the cross-correlation matrix are represented. Indicates the number of samples. Indicates the sample index. and This represents the features of two simulated 2D images derived from different stylistic versions of the same 3D wound image. and They represent the first time. In the samples, the The Dimensions and the stated The The value of the dimension;

[0077] Step S3322: Reduce redundancy, which reduces redundancy between features without increasing the differences in the feature structure. Specifically, it calculates the redundancy loss function using the following formula: ;

[0078] In the formula, Represents the redundancy loss function. Represents the cross-correlation matrix, This represents the diagonal elements of the cross-correlation matrix. and Both represent feature dimension indices. Represents the invariant term. Indicates the redundant weighting factor. Indicates the redundancy reduction term;

[0079] By minimizing the redundancy loss function, i.e., minimizing the off-diagonal elements of the cross-correlation matrix, redundancy between features is reduced;

[0080] Step S333: Feature dimensionality reduction; Principal component analysis is used to reduce the dimensionality of high-dimensional features;

[0081] Step S334: Style transfer, used to reduce registration errors caused by style differences. Specifically, through comparative learning, the gap between the simulated 2D image and the real 2D image is narrowed, and style transfer from the real 2D image style to the simulated 2D image style is achieved.

[0082] Example 4, see Figure 1 and Figure 4 This embodiment is based on the above embodiment. In step S4, the training phase specifically includes the following steps:

[0083] Step S41: Pre-training, using a simulated 2D image generated from the 3D wound image to pre-train the deep learning neural network, which learns how to register the 3D wound image with the simulated 2D image;

[0084] Step S42: Joint training, using simulated 2D images and real 2D images generated from 3D wound images to jointly train the deep learning neural network, which processes both simulated 2D images and real 2D images simultaneously;

[0085] Step S43: The trained deep learning neural network outputs a fused and registered image.

[0086] Example 5, see Figure 1 This embodiment is based on the above embodiment. In step S6, the statistical wound feature data specifically includes the following steps:

[0087] Step S61: Based on the fused registration image and the result of the first classification, the wound region is segmented from the fused registration image to obtain the temperature distribution of the wound region. Histogram equalization technology is used to enhance the contrast of the wound region. Machine vision technology is used to calculate the wound area. Based on the temperature distribution of the wound region in the fused registration image, temperature feature data is calculated. The temperature feature data includes the average temperature, median, standard deviation, skewness, and kurtosis of the wound region at each time point.

[0088] Step S62: Construct a mathematical model using partial differential equations. The mathematical model is used to quantify the effects of inflammatory cells and inflammatory mediators on the concentration of vascular endothelial growth factor and the density of newly formed capillary tips.

[0089] Step S63: Medication record, specifically, collecting the frequency of wound dressing and the usage of dressing products, and recording it as dressing information;

[0090] Step S64: Record the temperature characteristic data, the mathematical model, and the dressing information as wound characteristic data;

[0091] Step S65: Statistical analysis is used to screen wound feature data. Specifically, the Kruskal-Wallis test is used with a 95% confidence interval to analyze the correlation between the wound feature data and the wound healing status, and wound feature data related to the wound healing status are screened.

[0092] Example 6, see Figure 1This embodiment is based on the above embodiment. In step S5, the tissue type is specifically non-wound tissue, fibrin, granulation tissue, and necrotic tissue.

[0093] Example 7, see Figure 1 This embodiment is based on the above embodiment. All models and analyses in this solution are performed on a system equipped with an NVIDIA Titan RTX GPU. Python, Tensorflow, and scikit-learn software libraries are used for model training, parameter tuning, and performance evaluation.

[0094] Example 8, see Figures 1 to 3 This embodiment is based on the above embodiment. In step S3322, the redundancy weight factor is set to 0.005.

[0095] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0096] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0097] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for predicting wound healing time, characterized in that: The method includes the following steps: Step S1: Information acquisition, specifically, acquiring two-dimensional wound images through an RGB camera and an infrared thermal imaging camera, and using magnetic resonance imaging technology to obtain a three-dimensional MRI image containing soft tissue anatomical information, and recording the MRI three-dimensional image as a three-dimensional wound image; Step S2: Primary classification, specifically, using a deep neural network to perform binary classification on the two-dimensional wound image, dividing the two-dimensional wound image into wound region and non-wound region; Step S3: Information complementarity for image registration, ensuring that the structure of the 3D wound image can be accurately located in the 2D wound image. Specifically, deep learning technology is used to construct a deep learning neural network, which spatially aligns and fuses the 2D and 3D wound images, and outputs a fused and registered image. Step S4: Training phase; Step S5: Secondary classification for identifying tissue types. Specifically, based on the segmentation results of the primary classification, an image segmentation network based on a multilayer perceptron is used to segment the fused and registered image into fixed-size pixel blocks. Each pixel block is classified to identify the tissue type contained in each pixel block; prediction scores for different tissue types contained in each pixel block are calculated. Step S6: Collect wound feature data to extract and filter wound feature data; Step S7: Predict healing time. Specifically, the selected wound feature data and the predicted score are used as combined features. A Gaussian process regression model is used with the combined features as input data to predict the number of days required for wound healing.

2. The method for predicting wound healing time according to claim 1, characterized in that: In step S3, the information complementarity specifically includes the following steps: Step S31: Simulation image generation, used to generate simulated 2D images to expand the training dataset. Specifically, an encoder is used to generate a simulated 2D wound image corresponding to the 3D wound image. The simulated 2D wound image corresponding to the 3D wound image is recorded as the simulated 2D image, and the real 2D wound image corresponding to the 3D wound image is recorded as the real 2D image. Step S32: Reduce degrees of freedom, specifically by using a preset performance control technique as a constraint to limit the rotation and translation required for the 3D wound image to transform from the initial registration estimate to the optimal registration estimate. The initial registration estimate represents the initial position and orientation of the three-dimensional wound image; The optimal registration estimate represents the position and orientation when the three-dimensional wound image and the two-dimensional wound image are perfectly aligned, i.e., the final position and orientation; Step S33: Feature adaptation, used to reduce the difference between simulated 2D images and real 2D images at the feature level. Specifically, it involves using a generative adversarial network to reduce differences in feature structure and image style. The difference in feature structure specifically refers to the difference in feature content representation between the simulated 2D image and the real 2D image, and the difference in image style specifically refers to the difference in visual appearance between the simulated 2D image and the real 2D image. The three-dimensional wound image and the corresponding real 2D image are taken at different times, with different imaging devices used, and with different body positions of the injured person. These real 2D images are denoted as unpaired real images. A simulated two-dimensional wound image is generated based on the unpaired real images. The simulated two-dimensional wound image generated based on the unpaired real images is denoted as an unpaired simulated image. Using the unpaired real graph and the unpaired simulated graph, the deep learning neural network performs feature domain adaptation operations through unsupervised learning.

3. The method for predicting wound healing time according to claim 2, characterized in that: In step S33, the feature domain adaptive operation specifically includes the following steps: Step S331: Reduce feature structure differences to increase the realism of the simulated 2D image generated by the encoder. Specifically, calculate the encoder's adversarial loss function using the following formula: ； In the formula, This represents the adversarial loss function of the encoder. and These represent simulated 2D images and real 2D images, respectively. Indicates input data, This represents the expected value calculated by sampling feature maps from a simulated 2D image as input data. This represents the expected value calculated by sampling feature maps from a real 2D image as input data. Representing feature maps, Indicates the discriminator; The adversarial loss function is minimized through backpropagation; Step S332: Reduce image style differences; Step S333: Feature dimensionality reduction; Principal component analysis is used to reduce the dimensionality of high-dimensional features; Step S334: Style transfer, used to reduce registration errors caused by style differences. Specifically, through comparative learning, the gap between the simulated 2D image and the real 2D image is narrowed, and style transfer from the real 2D image style to the simulated 2D image style is achieved.

4. The method for predicting wound healing time according to claim 3, characterized in that: In step S332, reducing image style differences specifically includes the following steps: Step S3321: Calculate feature correlation to evaluate the linear relationship between different feature dimensions of the simulated 2D image. Calculate the cross-correlation matrix between features of different style versions of the simulated 2D image generated from the same 3D wound image. The calculation formula is as follows: ； In the formula, The elements of the cross-correlation matrix are represented. Indicates the number of samples. Indicates the sample index. and This represents the features of two simulated 2D images derived from different stylistic versions of the same 3D wound image. and They represent the first time. In the samples, the The Dimensions and the stated The The value of the dimension; Step S3322: Reduce redundancy, which reduces redundancy between features without increasing the differences in the feature structure. Specifically, it calculates the redundancy loss function using the following formula: ； In the formula, Represents the redundancy loss function. Represents the cross-correlation matrix, This represents the diagonal elements of the cross-correlation matrix. and Both represent feature dimension indices. Represents the invariant term. Indicates the redundant weighting factor. Indicates the redundancy reduction term; By minimizing the redundancy loss function, i.e., minimizing the off-diagonal elements of the cross-correlation matrix, redundancy between features is reduced.

5. The method for predicting wound healing time according to claim 4, characterized in that: In step S4, the training phase specifically includes the following steps: Step S41: Pre-training, using a simulated 2D image generated from the 3D wound image to pre-train the deep learning neural network, which learns how to register the 3D wound image with the simulated 2D image; Step S42: Joint training, using simulated 2D images and real 2D images generated from 3D wound images to jointly train the deep learning neural network, which processes both simulated 2D images and real 2D images simultaneously; Step S43: The trained deep learning neural network outputs a fused and registered image.

6. The method for predicting wound healing time according to claim 5, characterized in that: In step S6, the statistical wound feature data specifically includes the following steps: Step S61: Based on the fused registration image and the result of the first classification, the wound region is segmented from the fused registration image to obtain the temperature distribution of the wound region. Histogram equalization technology is used to enhance the contrast of the wound region. Machine vision technology is used to calculate the wound area. Based on the temperature distribution of the wound region in the fused registration image, temperature feature data is calculated. The temperature feature data includes the average temperature, median, standard deviation, skewness, and kurtosis of the wound region at each time point. Step S62: Construct a mathematical model using partial differential equations. The mathematical model is used to quantify the effects of inflammatory cells and inflammatory mediators on the concentration of vascular endothelial growth factor and the density of newly formed capillary tips. Step S63: Medication record, specifically, collecting the frequency of wound dressing and the usage of dressing products, and recording it as dressing information; Step S64: Record the temperature characteristic data, the mathematical model, and the dressing information as wound characteristic data; Step S65: Statistical analysis is used to screen wound feature data. Specifically, the Kruskal-Wallis test is used with a 95% confidence interval to analyze the correlation between the wound feature data and the wound healing status, and wound feature data related to the wound healing status are screened.

Images