Self-adaptive analysis nursing method based on wound image

Abstract

The invention discloses a self-adaptive analysis nursing method based on a wound image. The method comprises the following steps: S1, acquiring multi-modal image data of a wound; s2, performing fusion processing on the acquired multi-modal image data to generate a fused image; s3, analyzing the fused image by using a deep learning target detection model, identifying boundary, type, area and depth information of the wound, and forming a wound information set; s4, according to the identified wound information set, combining the physical condition data of the patient, and generating a model through a self-adaptive nursing strategy to formulate a personalized nursing scheme; s5, in the nursing process, the change condition of the wound is monitored in real time, the multi-modal image data are collected again at regular intervals, the steps S2 to S4 are repeated, and the nursing scheme is dynamically adjusted according to a new analysis result. According to the method, wound information is accurately obtained, a personalized nursing scheme is formulated, dynamic adjustment is achieved, and the accuracy, effectiveness and intelligent level of wound nursing are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical care, and particularly to an adaptive analysis nursing method based on wound images. Background Art

[0002] In the modern medical field, wound care plays a crucial role in the rehabilitation process and quality of life of patients. Accurately assessing the wound condition and formulating personalized nursing plans can effectively promote wound healing, reduce the risk of infection, and shorten the patient's hospital stay. However, traditional wound care methods are difficult to meet the clinical needs. On the one hand, traditional methods mainly rely on the subjective experience of medical staff, and there are differences in the judgment criteria of different medical staff, resulting in low accuracy and reliability of wound assessment; on the other hand, for complex wounds, it is difficult for traditional methods to comprehensively obtain various information of the wound, such as depth, tissue viability, etc., so that a precise and effective nursing plan cannot be formulated.

[0003] Currently, in the field of wound care technology, there are already some image-based analysis methods. These existing technologies use single-modal images, such as only using visible light images to observe the appearance of the wound. Although this method can obtain some surface information of the wound, such as color, shape, etc., it is impossible to deeply understand the internal physiological condition of the wound. For example, it is impossible to know the temperature distribution of the wound tissue and difficult to judge whether there is inflammation in the wound; nor can specific biomarkers be detected by fluorescence imaging to evaluate the potential process of wound healing. In addition, when formulating nursing plans, the existing technologies often do not fully consider the individual differences of patients, such as personal medical history, age, physical condition, etc., resulting in poor pertinence of the nursing plan and difficult to achieve the best nursing effect.

[0004] The adaptive analysis nursing method based on wound images of the present invention effectively solves the problems existing in the prior art. By obtaining multi-modal image data of the wound, fusing visible light images, infrared thermal imaging images, and fluorescence imaging images, it is possible to comprehensively and accurately obtain information about the wound, including boundaries, types, areas, and depths, etc. Combining the personal medical history, age, and physical condition data of the patient, using an adaptive nursing strategy generation model, a highly personalized nursing plan is formulated. During the nursing process, the wound changes are monitored in real time, and the nursing plan is dynamically adjusted to ensure the effectiveness and timeliness of nursing. Summary of the Invention

[0005] Based on the above content, the present application proposes an adaptive analysis nursing method based on wound images, including:

[0006] S1. Obtain multi-modal image data of the wound, where the multi-modal image data at least includes visible light images, infrared thermal imaging images, and fluorescence imaging images;

[0007] S2. Perform fusion processing on the acquired multi-modal image data to generate a fused image;

[0008] S3. Analyze the fused image using a deep learning object detection model to identify the boundary, type, area, and depth information of the wound, forming a wound information set;

[0009] S4. According to the identified wound information set, combined with the patient's personal medical history, age, and physical condition data, formulate a personalized care plan through an adaptive care strategy generation model;

[0010] S5. During the care process, continuously monitor the changes in the wound, regularly re-acquire multi-modal image data and repeat steps S2 - S4, and dynamically adjust the care plan according to the new analysis results.

[0011] Preferably, in step S1, acquiring the multi-modal image data of the wound specifically includes the following steps:

[0012] S11. Locate the wound, identify the color difference and texture difference features between the wound edge and the surrounding skin, determine the center position and range of the wound, and control the adjustable lens assembly to automatically adjust the focal length and shooting angle so that the wound is completely centered in the shooting field of view;

[0013] S12. Perform image acquisition to obtain a visible light image of the wound surface texture and color information, and acquire a temperature distribution image of the wound and its surrounding area;

[0014] S13. Apply a specific fluorescent reagent to the wound area, allow the reagent to fully react with the wound tissue, emit excitation light of the emission wavelength, filter out environmental light interference using a filter, and acquire a fluorescence emission image to complete the acquisition of multi-modal image data.

[0015] Preferably, in step S2, extract the features of each modal image through a feature extraction algorithm; the extraction of the features of the modal image through the feature extraction algorithm specifically includes visible light image feature extraction, infrared thermal imaging image feature extraction, and fluorescence imaging image feature extraction, obtaining the feature vector V of the visible light image, the feature vector IR of the infrared thermal imaging image, and the feature vector F of the fluorescence imaging image Y ; The extracted features are weighted and fused according to a preset weight to obtain a fused feature vector M, and then a fused image is reconstructed based on the fused feature vector M.

[0016] Preferably, the visible light image feature extraction is specifically as follows:

[0017] The visible light image feature extraction is performed by a convolutional neural network for feature extraction, setting the visible light image I v , through the convolutional layer C v1 , C v2, …, C vn Perform a convolution operation. At the i j th convolutional layer C vi The output feature map is Its calculation formula is Where K vi is the i j th convolutional kernel, * represents the convolution operation, is the bias term, ReLU is the activation function; the feature vector of the visible light image obtained after multi-layer convolution is

[0018] Preferably, the extraction of the infrared thermal imaging image features is specifically as follows:

[0019] Through two-dimensional discrete wavelet transform, wavelet decomposition of the image is performed to obtain sub-band coefficients of different frequencies. For the low-frequency sub-band coefficient A l Perform PCA dimensionality reduction processing to obtain the feature vector of the infrared thermal imaging image; when performing PCA dimensionality reduction processing on the low-frequency sub-band coefficient, calculate the covariance matrix, and the formula is: Where S y is the number of samples, is the i y th sample, is the sample mean. Perform eigenvalue decomposition on the covariance matrix C to obtain eigenvalues λ1 ≥ λ2 ≥ … ≥ λ q and the corresponding eigenvectors ±e1, e2, …, e q , select the eigenvectors corresponding to the first r largest eigenvalues to form the projection matrix P = [e1, e2, …, e r , project the low-frequency sub-band coefficient A l onto this projection matrix to obtain the dimensionality-reduced feature vector IR r , and the calculation formula is Form the feature vector IR of the infrared thermal imaging image.

[0020] Preferably, the extraction of the fluorescence imaging image features uses the local binary pattern (LBP) feature extraction algorithm to extract the texture and gray-scale distribution features of the image and generate the feature vector of the fluorescence imaging image, specifically as follows:

[0021] Taking each pixel point p(x, y) in the fluorescence imaging image I f as the center, select a circular neighborhood with a radius of R. The neighborhood contains sampling points. Perform bilinear interpolation on the sampling points in the neighborhood to obtain the gray value The gray value of the central pixel point is g c , calculate the LBP code value LBP R，N (x, y), and the formula is: Among them, s g (x) is a sign function, and its definition is: By traversing all pixel points in the fluorescence imaging image, the LBP feature map LBP map is obtained to acquire the texture features of the fluorescence imaging image;

[0022] The gray - level distribution features of the fluorescence imaging image are extracted through a gray - level co - occurrence matrix; the fluorescence imaging image I f constructs a gray - level co - occurrence matrix at a certain gray - level G, distance d, and direction θ. The number of occurrences of pixel pairs with gray - level values i h and j h is P(i h , j h ). Then, the element GLCM d，θ (i h , j h ) of the gray - level co - occurrence matrix is calculated by the formula: Texture features are obtained through the elements GLCM d，θ (i h , j h );

[0023] The LBP feature map LEP map and the texture features extracted from the gray - level co - occurrence matrix are combined to form the feature vector F Y of the fluorescence imaging image.

[0024] Preferably, the fusion image is reconstructed based on the fusion feature vector M, specifically:

[0025] The fusion feature vector M is input into an image reconstruction model improved by a generative adversarial network. The image reconstruction model consists of a generator G and a discriminator D; the generator G performs image reconstruction, and the feature map F G is output by the generator G, with a size of H×W×C, where H is the height, W is the width, and C is the number of channels. The feature map F G is processed by an adaptive weighted fusion layer. The weight matrix of the adaptive weighted fusion layer is W a , which is dynamically generated according to the feature importance of different regions in the feature map F G . For each pixel point (i x , j x ) in the feature map, its feature importance score S(i x , j x ) is calculated. The calculation formula is where α c is the importance coefficient of channel c, and the element w a (j x, j x ), the formula is where m represents the pixel index in the height direction of the feature map, n represents the pixel index in the width direction of the feature map, and the feature map F G is multiplied element-wise with the weight matrix W a and accumulated to obtain the reconstructed fused image I fusion (i x , j x ), and the calculation formula is

[0026] Preferably, in step S3, a deep learning object detection model is used to analyze the fused image to identify the boundary, type, area, and depth information of the wound, forming a wound information set, specifically including:

[0027] The deep learning object detection model is a model improved based on a convolutional neural network. In the model training stage, a large number of labeled multi-modal wound image data are used for training to optimize the model parameters;

[0028] The reconstructed fused image I fusion (i x , j x ) is input into the deep learning object detection model improved based on Transformer, and the multi-head self-attention mechanism is used to capture the global features to obtain F a , which is refined by a convolutional neural network to obtain F c , and the wound boundary is predicted through the formula P b (x j , y j ) = Sigmoid(W b F c (x j , y j ) + b b ), where W b is the boundary weight matrix, b b is the bias term, (x j , y j ) are the pixel coordinates, and the pixels greater than the threshold τ b are boundary points; F p is obtained through global average pooling, and the wound type is identified through the fully connected layer P t (i) = Softmax(W c F p + b c ). According to the number of pixels N b inside the wound boundary and the corresponding area S p of the pixels, the wound area is calculated by A = N b × S p , and combined with the multi-modal image features, through the formula Estimate the wound depth, where F d is the input feature, D(x, y) is the output wound depth value, and G k are different depth estimation sub-models, and w k is the corresponding weight coefficient, and b d is the bias term, and K is the number of sub-models;

[0029] Integrate the wound boundary, type, area, and depth information to form a wound information set.

[0030] Preferably, in step S4, according to the identified wound information set, combined with the patient's personal medical history, age, and physical condition data, a personalized care plan is formulated through an adaptive care strategy generation model, which specifically includes:

[0031] The adaptive care strategy generation model selects and combines the most suitable care plan from the care strategy templates based on the wound information and the patient's personal data using the decision tree algorithm;

[0032] Quantify the wound information set and the patient's personal data into feature vectors When constructing a decision tree, use the Gini impurity to measure the purity of the data set, where m is the number of categories, and p i is the proportion of samples belonging to the i-th category in the data set. Split the node according to selecting the feature and threshold with the minimum G s where D1 and D2 are the two sub-data sets after splitting, and |D|, |D1|, and |D2| represent the number of samples in the corresponding data sets respectively; The new feature vector After input, traverse the decision tree to the leaf node, and the node corresponds to the care strategy sub-template set {S1, S2,..., S k}, and calculate the matching degree score according to where w i is the weight of the feature x i , represents the matching degree between the feature x i and the corresponding feature conditions in the sub-template S j , and the value range is [0, 1]. By selecting the sub-template with a high score, a definite care plan is formed.

[0033] Preferably, in step S5, during the care process, collect wound data, transmit it to the HIS system through Bluetooth or 5G, use the LSTM model to predict the wound healing trend after data cleaning, trigger re-collection when abnormal, and then process and analyze the data according to S12 - S14 to generate a new care plan. According to the new plan, use the intelligent scheduling algorithm to allocate care resources to achieve dynamic adjustment of the care plan.

[0034] Compared with the prior art, the technical solution of the present application has the following technical effects:

[0035] The present invention solves the technical problem that traditional single-modal images cannot comprehensively obtain wound information by acquiring multi-modal image data of the wound (including visible light images, infrared thermal imaging images, and fluorescence imaging images) and performing fusion processing on them. Multi-modal images can reflect the wound condition from different angles. Visible light images present surface texture and color, infrared thermal imaging images show temperature distribution, and fluorescence imaging images detect specific biomarkers. After fusion processing, wound information can be comprehensively and accurately obtained, providing strong support for subsequent precise evaluation and nursing, obtaining more accurate wound assessment results, and improving the accuracy and pertinence of wound care.

[0036] The present invention uses a deep learning object detection model to analyze the fused image and formulates a personalized nursing plan in combination with the decision tree algorithm based on the wound information set and patient personal data, solving the problem that traditional nursing plans rely on subjective experience and lack pertinence. The deep learning model can accurately identify the wound boundary, type, area, and depth, and the decision tree algorithm selects the most suitable plan from the nursing strategy template. In this way, individual differences of patients can be fully considered, the nursing needs of different patients can be met, and the technical effects of improving the nursing effect, promoting faster wound healing, and enhancing the quality of patient recovery can be obtained.

[0037] The present invention monitors the wound changes in real time during the nursing process, regularly re-acquires multi-modal image data and repeats the analysis process to dynamically adjust the nursing plan, solving the problem that traditional nursing plans cannot respond to wound changes in time. Real-time monitoring and dynamic adjustment can optimize nursing measures in a timely manner according to different stages of wound healing, avoiding slow healing or increased infection risk caused by lagging nursing plans. Furthermore, the technical effects of ensuring that the nursing plan always conforms to the actual situation of the wound, effectively preventing complications, and improving nursing efficiency can be obtained.

[0038] The present invention reconstructs and generates a fused image through an image reconstruction model improved by a generative adversarial network based on the fused feature vector, solving the problem that traditional image fusion methods are difficult to highlight the key features of different modal images and have poor fusion effects. The generator and discriminator of the generative adversarial network play against each other, can better capture the features of each modal image, and the adaptive weighted fusion layer dynamically generates a weight matrix according to the feature importance. This enables the reconstructed fused image to retain the key information of the multi-modal images, and the features of different regions are clearer, obtaining the technical effects of improving the image quality, enhancing the image detail expressiveness, and providing a better image basis for subsequent wound feature recognition and analysis.

[0039] The above description is only an overview of the technical solution of the present application. In order to better understand the technical means of the present application, it can be implemented according to the content of the specification. In addition, in order to make the above and other purposes, features, and advantages of the present application more obvious and understandable, the following describes the preferred embodiments of the present application in detail in conjunction with the accompanying drawings.

[0040] From the following detailed description of the specific embodiments of the present application in conjunction with the accompanying drawings, those skilled in the art will more clearly understand the above and other purposes, advantages, and features of the present application. Brief Description of the Drawings

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following briefly introduces the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings. In all the drawings, similar elements or parts are generally denoted by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0042] Figure 1 Flowchart of the adaptive analysis and nursing method based on wound images of the present invention;

[0043] Figure 2 Flowchart of obtaining multi-modal image data of a wound in the adaptive analysis and nursing method based on wound images of the present invention;

[0044] Figure 3 Flowchart of the experimental model structure of the adaptive analysis and nursing method based on wound images of the present invention;

[0045] Figure 4 Experimental verification comparison chart of the adaptive analysis and nursing method based on wound images of the present invention. Detailed Description of the Specific Embodiments

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the following clearly and completely describes the technical solutions in the embodiments of the present application in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are some, but not all, of the embodiments of the present application. In the following description, specific details such as specific configurations and components are provided only to assist in a comprehensive understanding of the embodiments of the present application. Therefore, those skilled in the art should clearly understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present application. In addition, descriptions of known functions and structures are omitted for clarity and conciseness.

[0047] It should be understood that the "one embodiment" or "this embodiment" mentioned throughout the specification means that the specific features, structures or characteristics related to the embodiment are included in at least one embodiment of the present application. Therefore, the "one embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. In addition, these specific features, structures or characteristics can be combined in one or more embodiments in any suitable manner.

[0048] In addition, this application may repeat reference numerals and / or letters in different instances. This repetition is for the purpose of simplicity and clarity and does not in itself indicate the relationship between the various embodiments and / or arrangements discussed.

[0049] The term "and / or" in this article is merely a description of the relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another relationship between associated objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.

[0050] The term "at least one" in this article is merely a description of the relationship between associated objects, indicating that three relationships can exist. For example, at least one of A and B can mean: A exists alone, A and B exist simultaneously, and B exists alone.

[0051] It should also be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion.

[0052] Embodiment 1

[0053] This embodiment details an adaptive analysis and care method based on wound images, as Figure 1 shown, including:

[0054] S1. Obtain multi-modal image data of the wound, and the multi-modal image data at least includes visible light images, infrared thermal imaging images, and fluorescence imaging images;

[0055] S2. Perform fusion processing on the obtained multi-modal image data to generate a fused image;

[0056] S3. Analyze the fused image using a deep learning object detection model to identify the boundary, type, area, and depth information of the wound, forming a wound information set;

[0057] S4. According to the identified wound information set, combined with the patient's personal medical history, age, and physical condition data, formulate a personalized care plan through an adaptive care strategy generation model;

[0058] S5. During the nursing process, monitor the changes of the wound in real time, regularly re-collect multi-modal image data and repeat steps S2 - S4, and dynamically adjust the care plan according to the new analysis results.

[0059] Furthermore, as Figure 2 shown, obtaining the multi-modal image data of the wound in step S1 specifically includes the following steps:

[0060] S11. Locate the wound, identify the color difference and texture difference features between the wound edge and the surrounding skin, determine the center position and range of the wound, and control the adjustable lens assembly to automatically adjust the focal length and shooting angle so that the wound is completely centered in the shooting field of view;

[0061] S12. Perform image acquisition to obtain a visible light image of the wound surface texture and color information, and acquire a temperature distribution image of the wound and its surrounding area;

[0062] S13. Apply a specific fluorescent reagent to the wound area, make the reagent fully react with the wound tissue, emit excitation light of a specific wavelength, use a filter to filter out environmental light interference, and acquire a fluorescence emission image to complete the acquisition of multi-modal image data.

[0063] Furthermore, in step S2, the features of each modal image are extracted respectively through a feature extraction algorithm; extracting the features of the modal image through a feature extraction algorithm specifically includes visible light image feature extraction, infrared thermal imaging image feature extraction, and fluorescence imaging image feature extraction, obtaining the feature vector V of the visible light image, the feature vector IR of the infrared thermal imaging image, and the feature vector F of the fluorescence imaging image Y ; The extracted features are weighted and fused according to a preset weight to obtain a fused feature vector M, and then a fused image is reconstructed based on the fused feature vector M.

[0064] Furthermore, the visible light image feature extraction is specifically as follows:

[0065] The visible light image feature extraction is performed through a convolutional neural network to extract features, setting the visible light image I v , through the convolutional layer C v1 , C v2 , …, C vn perform convolutional operations, and in the i j th convolutional layer Cvi The output feature map is Its calculation formula is where K vi is the i-th j layer convolution kernel, * represents the convolution operation, is the bias term, ReLU is the activation function; after multiple convolutions, the feature vector of the visible light image is obtained

[0066] Furthermore, the feature extraction of the infrared thermal imaging image is specifically as follows:

[0067] Through two-dimensional discrete wavelet transform, the image is wavelet decomposed to obtain different frequency sub-band coefficients. For the low-frequency sub-band coefficient A l PCA dimensionality reduction processing is performed to obtain the feature vector of the infrared thermal imaging image; when performing PCA dimensionality reduction processing on the low-frequency sub-band coefficients, the covariance matrix is calculated, and the formula is: where S y is the number of samples, is the i-th y sample, is the sample mean. The covariance matrix C is eigen-decomposed to obtain eigenvalues λ1 ≥ λ2 ≥ … ≥ λ q and the corresponding eigenvectors ±e1, e2, …, e q , and the eigenvectors corresponding to the first r largest eigenvalues are selected to form the projection matrix P = [e1, e2, …, e r , and the low-frequency sub-band coefficient A l is projected onto this projection matrix to obtain the dimensionality-reduced feature vector IR r , and the calculation formula is to form the feature vector IR of the infrared thermal imaging image.

[0068] Furthermore, the feature extraction of the fluorescence imaging image uses the local binary pattern (LBP) feature extraction algorithm to extract the texture and gray-scale distribution features of the image and generate the feature vector of the fluorescence imaging image, specifically as follows:

[0069] Taking each pixel point p(x, y) in the fluorescence imaging image I f as the center, a circular neighborhood with a radius of R is selected. The neighborhood contains sampling points, and bilinear interpolation is performed on the sampling points in the neighborhood to obtain the gray value The gray value of the central pixel point is g c , and the LBP coding value LBP R，N (x, y) is calculated, and the formula is: where s g (x) is the sign function, and its definition is: By traversing all pixel points in the fluorescence imaging image, the LBP feature map LBP is obtained map , and the texture features of the fluorescence imaging image are acquired;

[0070] The gray distribution features of the fluorescence imaging image are extracted through the gray-level co-occurrence matrix; the fluorescence imaging image I f Construct the gray-level co-occurrence matrix at a certain gray level G, distance d, and direction θ. For the pixel pair with gray values i h and j h in the image, the number of occurrences is P(i h , j h ). Then, the element GLCM of the gray-level co-occurrence matrix d，θ (i h , j h ) is calculated as follows: The texture features are obtained through the element GLCM of the gray-level co-occurrence matrix d，θ (i h , j h );

[0071] Combine the LBP feature map LBP map and the texture features extracted from the gray-level co-occurrence matrix to form the feature vector F of the fluorescence imaging image Y .

[0072] Furthermore, based on the fused feature vector M, a fused image is reconstructed as follows:

[0073] Input the fused feature vector M into the image reconstruction model improved by the generative adversarial network. The image reconstruction model consists of a generator G and a discriminator D; the generator G performs image reconstruction, and the feature map F G is output by the generator G, with a size of H×W×C, where H is the height, W is the width, and C is the number of channels. The feature map F G is processed by the adaptive weighted fusion layer. The weight matrix of the adaptive weighted fusion layer is W a , which is dynamically generated according to the feature importance of different regions in the feature map F G . For each pixel point (i x , j x ) in the feature map, calculate its feature importance score S(i x , j x ), and the calculation formula is where α c is the importance coefficient of channel c, and the element w a (i x , j x ) of the weight matrix is generated using the Softmax function according to the score. The formula is where m represents the pixel index in the height direction of the feature map, and n represents the pixel index in the width direction of the feature map. Multiply the feature map F G element-wise with the weight matrix W a and accumulate to obtain the reconstructed fused image I fusion (i x , j x ). The calculation formula is

[0074] Furthermore, in step S3, a deep learning object detection model is used to analyze the fused image to identify the boundary, type, area, and depth information of the wound, forming a wound information set, specifically including:

[0075] The deep learning object detection model is a model improved based on a convolutional neural network. In the model training stage, a large number of labeled multi-modal wound image data are used for training to optimize the model parameters;

[0076] Input the reconstructed fused image I fusion (i x , j x ) into the deep learning object detection model improved based on Transformer. Use the multi-head self-attention mechanism to capture global features to obtain F a , refine it through a convolutional neural network to obtain F c . Predict the wound boundary through the formula P b (x j , y j ) = Sigmoid(W b F c (x j , y j ) + b b ), where W b is the boundary weight matrix, b b is the bias term, (x j , y j ) are the pixel coordinates, and the pixels greater than the threshold τ b are boundary points; Obtain F p through global average pooling, and identify the wound type through the fully connected layer P t (i) = Softmax(W c F p + b c ). Calculate the wound area according to the number of pixels N b inside the wound boundary and the corresponding area S p of the pixels. From A = N b × S p , calculate the wound area, and combine multi-modal image features through the formula Estimate the wound depth, where F d is the input feature, D(x, y) is the output wound depth value, and G k are different depth estimation sub-models, and w k is the corresponding weight coefficient, and b d is the bias term, and K is the number of sub-models;

[0077] Integrate the wound boundary, type, area, and depth information to form a wound information set.

[0078] Furthermore, in S4, according to the identified wound information set, combined with the patient's personal medical history, age, and physical condition data, a personalized care plan is formulated through an adaptive care strategy generation model, which specifically includes:

[0079] The adaptive care strategy generation model selects and combines the most suitable care plan from the care strategy templates based on the wound information and the patient's personal data using the decision tree algorithm;

[0080] Quantify the wound information set and the patient's personal data into feature vectors When constructing the decision tree, use the Gini impurity to measure the purity of the data set, where m is the number of classes, and p i is the proportion of samples belonging to the i-th class in the data set, and split the node by selecting the feature and threshold with the minimum G where D1 and D2 are the two sub-data sets after splitting, and |D|, |D1|, and |D2| represent the number of samples in the corresponding data sets respectively; the new feature vector s After input, traverse the decision tree to the leaf node, and the node corresponds to the care strategy sub-template set {S1, S2,..., S }, and calculate the matching degree score according to k where w is the weight of the feature x i and i represents the matching degree between the feature x and the corresponding feature conditions in the sub-template S i , with a value range of [0, 1]. By selecting the sub-template with a high score, a definite care plan is formed. j

[0081] Furthermore, in step S5, during the nursing process, collect the wound data, transmit it to the HIS system via Bluetooth or 5G, predict the wound healing trend using the LSTM model after data cleaning, trigger re-collection when abnormal, and then process and analyze the data according to S12 - S14 to generate a new care plan. According to the new plan, use the intelligent scheduling algorithm to allocate nursing resources to achieve dynamic adjustment of the care plan

[0082] ​In this embodiment, by acquiring multi-modal image data such as visible light, infrared thermal imaging, and fluorescence imaging of the wound, the problem of limited information obtained by traditional single images is solved. Feature extraction algorithms are used to extract features from each modal image respectively and perform weighted fusion to generate a fused image, overcoming the defect of low accuracy in traditional image fusion. A deep learning object detection model improved based on convolutional neural network and Transformer is used to analyze the fused image, accurately identifying the wound boundary, type, area, and depth information. Compared with the traditional method relying on the subjective judgment of medical staff, the accuracy and objectivity of wound assessment are greatly improved. Combining the patient's personal medical history, age, and physical condition, an adaptive nursing strategy generation model is used to develop a personalized nursing plan, changing the drawback of the traditional "one-size-fits-all" nursing plan and making the nursing more targeted, effectively improving the nursing effect.

[0083] Based on Embodiment 1, after obtaining the fused feature vector in this embodiment, the process of reconstructing the fused image based on the fused feature vector M is described as follows:

[0084] Input the fused feature vector M into an image reconstruction model improved by a generative adversarial network. The image reconstruction model consists of a generator G and a discriminator D. The generator G performs image reconstruction. The generator G contains L transposed convolutional layers T1, T2,..., T L , and the input of the k-th transposed convolutional layer is Z k , and the output is Z k+1 , and its calculation formula is Z k+1 = σ(T k (Z k ) + b k ), where σ is the activation function, T k is the k-th transposed convolutional kernel, and b k is the bias term; the initial input is the fused feature vector M;

[0085] Output the feature map F G through the generator G, with the size of H×W×C, where H is the height, W is the width, and C is the number of channels. Process the feature map F G through an adaptive weighted fusion layer. The weight matrix of the adaptive weighted fusion layer is W a , which is dynamically generated according to the feature importance of different regions in the feature map F G . For each pixel point (i x , j x ) in the feature map, calculate its feature importance score S(i x , j x ), and the calculation formula is where α c is the importance coefficient of channel c, and the elements w a (ix , j x ), the formula is where m represents the pixel index in the height direction of the feature map, n represents the pixel index in the width direction of the feature map, and the feature map F G is multiplied element-wise with the weight matrix W a and accumulated to obtain the reconstructed fused image I fusion (i x , j x ), and the calculation formula is

[0086] In this embodiment, by inputting the fused feature vector into the image reconstruction model improved based on the generative adversarial network, the transposed convolutional layer of the generator is used to gradually restore the spatial structure of the image, so that the fused image can highly restore the key details of each modality image. The adaptive weighted fusion layer dynamically generates weights according to the feature importance, and can accurately highlight the key features of the wound. For example, in fluorescence imaging, it can more clearly show the boundary of the diseased tissue, and in infrared thermal imaging, it highlights the abnormal temperature area, greatly improving the quality and accuracy of the fused image, providing a better image basis for the subsequent accurate analysis of the wound, and effectively enhancing the reliability and comprehensiveness of the judgment of the wound condition.

[0087] Based on Embodiment 1, after obtaining the fused feature vector in this embodiment, the process of reconstructing and generating the fused image based on the fused feature vector M is described as follows:

[0088] In step S3, a deep learning object detection model is used to analyze the fused image to identify the boundary, type, area, and depth information of the wound, forming a wound information set, specifically including:

[0089] The deep learning object detection model is a model improved based on the convolutional neural network. In the model training stage, a large number of labeled multi-modal wound image data are used for training to optimize the model parameters;

[0090] The reconstructed fused image I fusion (i x , j x ) is input into the deep learning object detection model improved based on Transformer. The model captures the global features of the fused image through the multi-head self-attention mechanism. The input image is I, and the feature representation F is obtained after the multi-head self-attention calculation a , for the h-th attention head, the calculation process is:

[0091]

[0092] where is the linear transformation matrix of different attention heads, d kis the dimension of the key vector, H is the number of attention heads, A h The weighted feature matrix calculated by the h-th attention head;

[0093] Use the multi-head self-attention mechanism to capture global features to obtain F a , refined by a convolutional neural network to obtain F c , through the formula P b (x j , y j ) = Sigmoid(W b F c (x j , y j ) + b b ) to predict the wound boundary, where W b is the boundary weight matrix, b b is the bias term, (x j , y j ) are the pixel coordinates, and the pixels greater than the threshold τ b are the boundary points, thereby determining the wound boundary; through global average pooling to obtain F p , through the fully connected layer P t (i) = Softmax(W c F p + b c ) to identify the wound type, according to the number of pixels N b within the wound boundary and the corresponding area S p of the pixels, from A = N b × S p calculate the wound area, combine multi-modal image features, and through the formula estimate the wound depth, where F d is the input feature, D(x, y) outputs the wound depth value, G k are different depth estimation sub-models, w k are the corresponding weight coefficients, b d is the bias term, and K is the number of sub-models;

[0094] Integrate the wound boundary, type, area and depth information to form a wound information set.

[0095] In this embodiment, the technology of using a deep learning object detection model to analyze the fused image realizes the accurate recognition and quantification of various aspects of wound information, strongly promoting the intelligent and scientific process of wound care. By integrating the Transformer architecture and convolutional neural network, the model can efficiently capture the global and local features of the fused image, accurately locate the wound boundary, providing an accurate basis for evaluating the wound area; accurately identify the wound type, which helps nurses quickly judge the nature of the wound and take targeted nursing measures; the wound area calculated based on the boundary can objectively measure the change in the wound size and intuitively reflect the healing progress; the wound depth estimated by combining multi-modal image features further supplements the three-dimensional information of the wound, enabling nurses to comprehensively understand the wound condition. The finally formed wound information set provides a rich and accurate data basis for the formulation and adjustment of subsequent nursing plans, greatly improving the accuracy, effectiveness, and pertinence of wound care.

[0096] Based on Embodiment 1, this embodiment details the experimental study on the adaptive analysis nursing method based on wound images of this application, specifically as follows:

[0097] Select 100 wound patients as experimental subjects, covering different ages, genders, wound types, and physical conditions. Among them, there are 30 burn patients, 35 traumatic wound patients, 25 postoperative wound patients, and 10 chronic ulcer patients, and the wounds are in different healing stages; details of each patient's personal medical history, age, physical condition, etc. are recorded. For example, Patient A, 56 years old, male, has a history of diabetes, and the fasting blood glucose value is 6.23 mmol / L. This time, he was admitted to the hospital due to burns; Patient B, 32 years old, female, is in good health and has a wound due to accidental trauma.

[0098] Randomly divide the 100 patients into two groups. There are 50 patients in the experimental group, who adopt the adaptive analysis nursing method based on wound images of this application; there are 50 patients in the control group, who adopt the existing conventional nursing method.

[0099] Utilize the intelligent recognition function of the image acquisition device to locate the patient's wound. Taking Patient A as an example, by analyzing the color difference between the wound edge and the surrounding skin (such as the RGB color differences are R: 25.36, G: 18.45, B: 12.67) and the texture difference (the texture complexity index is 0.735), it is determined that the center position of the wound is at the image coordinates (256.34, 321.56), and the range radius is 45.67 mm. The adjustable lens assembly of the device automatically adjusts the focal length to 78.45 mm, and the shooting angle is 32.56°, ensuring that the wound is completely centered in the shooting field of view.

[0100] Obtain a visible light image. In the visible light image of Patient A, the surface texture roughness of the wound is Ra = 3.25 μm. In the HSV color space, the hue H is 0.053, the saturation S is 0.456, and the value V is 0.678. At the same time, collect an infrared thermal imaging image. The temperature at the center of the wound is 37.86 °C, the temperature of the surrounding normal skin is 36.54 °C, and the temperature difference is 1.32 °C.

[0101] Apply a specific fluorescent reagent to the wound area of Patient A. After sufficient reaction, use an excitation light with an emission wavelength of 490.23 nm, and use a filter to remove ambient light interference, and collect a fluorescence emission image. The maximum fluorescence intensity in the image is 456.78 (unit: fluorescence intensity unit), the minimum is 125.36, and the average fluorescence intensity is 256.34.

[0102] Perform feature extraction. Use a convolutional neural network to extract features from the visible light image. Set the visible light image I v , through multiple convolutional layers C v1 , C v2 , C v3 (In this experiment, three layers are taken as an example) convolutional operation. In the i j = 1 convolutional layer C v1 , the convolutional kernel K v1 has a size of 3×3, and its parameters are [0.147, 0.215, -0.082; 0.324, 0.103, 0.487; -0.201, 0.532, 0.115], and the bias term b v1 = 0.089. The output feature map F v1 = ReLU(K v1 * F v0 + b v1 ). After three layers of convolution, the feature vector of the visible light image is V = [0.268, 0.432, 0.157, -0.063, 0.314];

[0103] Perform two-dimensional discrete wavelet transform on the infrared thermal imaging image to obtain different frequency sub-band coefficients. Perform PCA dimensionality reduction processing on the low-frequency sub-band coefficients A l . The number of samples S y = 100. Calculate the covariance matrix C. For example, for the i y = 5th sample Sample mean The covariance matrix C is calculated as [0.048, 0.027, -0.015; 0.027, 0.038, 0.011; -0.015, 0.011, 0.042]; perform eigenvalue decomposition on the covariance matrix C to obtain eigenvalues λ1 = 0.062 ≥ λ2 = 0.045 ≥ λ3 = 0.022 and corresponding eigenvectors e1 = [0.765, 0.482, -0.436], e2 = [0.378, 0.843, 0.385], e3 = [-0.543, 0.187, 0.812]. Select the eigenvectors corresponding to the first r = 2 largest eigenvalues to form the projection matrix P = [0.765, 0.482; 0.378, 0.843], and project the low-frequency subband coefficients A l onto this projection matrix to obtain the feature vector IR after dimensionality reduction r = [0.256, 0.523], forming the feature vector IR of the infrared thermal imaging image.

[0104] Using the pixel point p(100.23, 150.45) in the fluorescence imaging image I f as the center, select a circular neighborhood with a radius R = 3.23. The neighborhood contains N = 8 sampling points. Perform bilinear interpolation on the sampling points to obtain the gray values such as g1 = 125.36, g2 = 130.45,..., and the gray value of the central pixel point g c == 128.56, and calculate the LBP coding value which is calculated as 10101010 (binary, converted to decimal is 170). By traversing all pixel points, obtain the LBP feature map LBP map ; at the same time, construct a gray-level co-occurrence matrix at the gray level G = 256, distance d = 2.34, and direction θ == 45°, and calculate the elements GLCM of the gray-level co-occurrence matrix d，θ (i h , j h ), such as the number of occurrences P(105, 125) = 62 of the pixel pair with gray values i h = 105 and j h = 125, then combine the LBP feature map LBP map and the texture features extracted from the gray-level co-occurrence matrix to form the feature vector F of the fluorescence imaging image Y .

[0105] Weightedly fuse the extracted features according to the preset weights (weight of the visible light image feature vector is 0.4, weight of the infrared thermal imaging image feature vector is 0.3, weight of the fluorescence imaging image feature vector is 0.3) to obtain the fused feature vector M. Input M into the image reconstruction model improved by the generative adversarial network, and the generator G outputs the feature map FG , with a size of 256×256×3; process F through an adaptive weighted fusion layer G , for the pixel point (i x , j x ) = (150.34, 200.56), calculate its feature importance score Let α1 == 0.22, α2 == 0.33, α3 == 0.45, F = (150.34, 200.56, 1) = 0.185, F = (150.34, 200.56, 2) = 0.276, F = (150.34, 200.56, 3) = 0.358, then S(150.34, 200.56) = 0.22×0.185 + 0.33×0.276 + 0.45×0.358 = 0.298. Generate the elements of the weight matrix using the Softmax function Calculated to be 0.015, multiply the feature map F G element-wise with the weight matrix w a and accumulate to obtain the reconstructed fused image I fusion (150.34, 200.56).

[0106] As Figure 3 shown, use the improved deep learning object detection model based on the convolutional neural network to analyze the fused image. In the model training stage, use 8000 labeled multi-modal wound image data for training to optimize the model parameters. Input the reconstructed fused image I fusion into the improved deep learning object detection model based on Transformer, use the multi-head self-attention mechanism to capture global features to obtain F a , and refine it through the convolutional neural network to obtain F c . When predicting the wound boundary, through the formula P b (x j , y j ) = Sigmoid(w b F c (x j , y j ) + b b ), let the boundary weight matrix W b be [0.156, 0.224; 0.305, 0.412], the bias term b b = 0.062, for the pixel point (x j , y j ) = (200.34, 250.67), F c (200.34, 250.67) = [0.256, 0.314], then P b(200.34, 250.67) = Sigmoid ([0.156, 0.224; 0.305, 0.412] × [0.256, 0.314] + 0.062) = 0.684, greater than the threshold τ b = = 0.5, the pixel is a boundary point. Through global average pooling, we get F p , through the fully connected layer P t (i) = Softmax(W c F p +b c ) Identify the wound type and identify the wound type of patient A as a deep second-degree burn; according to the number of pixels N within the wound boundary b =12568.34 and the corresponding area of ​​the pixel S p = 0.0125 square millimeters, and the wound area A = 157.1 square millimeters. Combining multimodal image features, the formula Estimate wound depth. Let K = 3, w1 = 0.28, w2 = 0.42, w3 = 0.30, G1, G2, G3 are different depth estimation sub-models, F d (200.34, 250.67)=[0.152, 0.243, 0.315], b d =0.05, and the wound depth is calculated to be 2.45 mm. The wound boundary, type, area and depth information are integrated to form a wound information set.

[0107] According to the wound information set of patient A (deep second-degree burn, area 157.1 square millimeters, depth 2.45 millimeters), combined with his personal medical history (diabetes history), age (56 years old), and physical condition (fasting blood glucose 6.23mmol / L), a personalized nursing plan is formulated through the adaptive nursing strategy generation model. The wound information set and the patient's personal data are quantified into feature vectors X. When constructing the decision tree, the Gini impurity is used To measure the purity of the data set, let the number of categories m = 5 (different nursing plan categories), the proportion of samples belonging to category 1 in the data set is p1, category 2 is p2, category 3 is p3, category 4 is p4, category 5 is p5, p1 = 0.21, p2 = 0.15, p3 = 0.32, p4 = 0.23, p5 = 0.09, then G(D) = 1-(0.21 2 +0.15 2 +0.32 2 +0.23 2 +0.09 2 )=0.805. Select the smallest G s The new feature vector X is input and then traverses the decision tree to the leaf node, which corresponds to the nursing strategy sub-template set S1, S2, x3. Calculate the matching degree score. The weight of feature x1, w1 = 0.35, and the weight of feature x2, w2 = 0.65. For sub-template S1, then Score1 = 0.35×0.75 + 0.65×0.80 = 0.7825. By selecting the sub-template with the higher score, the nursing plan developed for patient A includes: covering the wound with a silver ion-containing antibacterial dressing every day and changing it every 3 days; controlling the patient's blood sugar level and maintaining the fasting blood sugar at 5.0 - 6.0 mmol / L; performing wound debridement regularly, flushing the wound with normal saline, and controlling the debridement time at 15 - 20 minutes each time.

[0108] During the nursing process, a wearable sensor is used to collect real-time data such as the temperature and humidity of patient A's wound, and the data is transmitted to the HIS system via Bluetooth. Multimodal image data is collected every 3 days. After data cleaning, an LSTM model is used to predict the wound healing trend. On the 7th day of nursing, it is predicted that the wound healing speed of patient A is slower than expected, triggering the re-collection of multimodal image data. After the re-collected data is processed and analyzed according to the above steps, it is found that there are slight signs of wound infection. According to the new analysis results, the nursing plan is adjusted as follows: increasing the replacement frequency of the antibacterial dressing to 2 times a day; using an antibiotic ointment locally on the wound; strengthening blood sugar monitoring and strictly controlling the fasting blood sugar at 5.0 - 5.5 mmol / L. At the same time, an intelligent scheduling algorithm is used to allocate nursing resources, such as arranging experienced nurses to be responsible for the nursing of this patient.

[0109] Fifty patients in the control group adopted the existing conventional nursing method, that is, relying only on the experience of nurses and simple wound observation for nursing. In the judgment of wound type, there is a relatively high misdiagnosis rate. For example, some deep second-degree burns are misdiagnosed as superficial second-degree burns, and the misdiagnosis rate reaches 20%. In terms of wound area measurement, simple measurement tools and estimation methods are used, and the measurement error is relatively large, with an average error reaching 35%. For the wound depth, it is basically impossible to accurately judge, and only a rough classification of depth can be made, as shown in the following table:

[0110]

[0111] In the formulation of the nursing plan, a unified nursing process is adopted without considering the individual differences of patients. For example, for patients with a history of diabetes, there are no targeted nursing measures to adjust the impact of blood sugar control on wound healing. During the nursing process, there is a lack of scientific monitoring of the wound healing trend, and the nursing plan is only adjusted when the wound shows obvious deterioration, resulting in an extended wound healing time. The average healing time is 5.23 days more than that of the experimental group, and the infection incidence rate reaches 16%, which is higher than 6% of the experimental group, as shown in the following table:

[0112] Comparison items Technical solution of the present application Existing technology Average wound healing time (days) 14.56 19.79 Incidence of wound infection 6％ 16％

[0113] It can be clearly seen from the above table comparison and the detailed experimental process that the adaptive analysis nursing method based on wound images in this application is significantly superior to the prior art in terms of the accuracy of wound information recognition and the nursing effect.

[0114] As Figure 4 shown, by comparing patients with similar wounds on burn wounds in the experimental group and the control group, it is found that the curve obtained through the experiment of this application is curve B, while the curve obtained through ordinary nursing is curve D; at the same time, by comparing patients with similar wounds on wounds caused by accidental trauma in the experimental group and the control group, it is found that the curve obtained through the experiment of this application is curve A, while the curve obtained through ordinary nursing is curve C. Therefore, the superiority of this application is further verified.

[0115] This embodiment verifies through experiments that this application can perform wound nursing more scientifically and accurately through multi-modal image fusion, deep learning model analysis, and the formulation and dynamic adjustment of personalized nursing plans, effectively improving the nursing quality and reducing the pain and recovery time of patients.

[0116] The above are only the preferred embodiments of the present invention, and it does not limit the protection scope of the present invention accordingly. For those skilled in the art, the present invention can have various changes and modifications; within the spirit and principle of the present invention, any changes, modifications, substitutions, integrations, and parameter changes to these embodiments by means of conventional substitutions or the ability to achieve the same function without departing from the principle and spirit of the present invention fall within the protection scope of the present invention.

Claims

1. An adaptive wound image-based nursing method, characterized in that: include: S1. Acquire multimodal image data of a wound, wherein the multimodal image data at least includes a visible light image, an infrared thermal imaging image, and a fluorescent imaging image; S2, performing fusion processing on the acquired multimodal image data to generate a fused image; S3. Use the deep learning target detection model to analyze the fused image, identify the boundary, type, area and depth information of the wound, and form a wound information set; S4. Based on the identified wound information set, combined with the patient’s personal medical history, age, and physical condition data, a personalized nursing plan is developed through an adaptive nursing strategy generation model; S5. During the nursing process, monitor the changes of the wound in real time, regularly re-collect multimodal image data and repeat steps S2-S4, and dynamically adjust the nursing plan according to the new analysis results.

2. The wound image-based adaptive analysis nursing method according to claim 1, characterized in that: Acquiring multimodal image data of the wound in step S1 specifically includes the following steps: S11, locate the wound, identify the color difference and texture difference characteristics between the wound edge and the surrounding skin, determine the center position and range of the wound, and control the adjustable lens assembly to automatically adjust the focal length and shooting angle so that the wound is completely in the center of the shooting field of view; S12, performing image acquisition to obtain a visible light image of the wound surface texture and color information, and an image of the temperature distribution of the wound and its surrounding area; S13. Apply a specific fluorescent reagent to the wound site to allow the reagent to fully react with the wound tissue, emit excitation light of the wavelength, use a filter to filter out ambient light interference, collect fluorescence emission images, and complete the acquisition of multimodal image data.

3. The wound image-based adaptive analysis nursing method according to claim 1, characterized in that: The step S2 extracts the features of each modal image respectively by a feature extraction algorithm; the feature extraction of the modal image by the feature extraction algorithm specifically includes visible light image feature extraction, infrared thermal imaging image feature extraction and fluorescence imaging image feature extraction, and obtains a feature vector V of the visible light image, a feature vector IR of the infrared thermal imaging image, and a feature vector F of the fluorescence imaging image. Y The extracted features are weightedly fused according to preset weights to obtain a fused feature vector M, and then a fused image is reconstructed based on the fused feature vector M.

4. The wound image-based adaptive analysis nursing method according to claim 2 or 3, characterized in that: The visible light image feature extraction is specifically as follows: The visible light image feature extraction is performed by a convolutional neural network. v , through the convolutional layer C v1 , C v2 , C vn Perform convolution operation on the i-th j Convolutional layer C vi The output feature map is The calculation formula is: Where K vi is the i j Layer convolution kernel, * indicates convolution operation, is the bias term, ReLU is the activation function; the feature vector of the visible light image is obtained after multi-layer convolution 5. The wound image-based adaptive analysis nursing method according to claim 2 or 3, characterized in that: The infrared thermal imaging image feature extraction is specifically as follows: Through two-dimensional discrete wavelet transform, the image is decomposed into wavelet to obtain different frequency sub-band coefficients, and the low-frequency sub-band coefficient A l Perform PCA dimensionality reduction processing to obtain the feature vector of the infrared thermal imaging image; when the low-frequency sub-band coefficients are subjected to PCA dimensionality reduction processing, the covariance matrix is ​​calculated, and the formula is: Where S y is the sample size, For the i y samples, is the sample mean, and the covariance matrix C is decomposed into eigenvalues ​​to obtain eigenvalues ​​λ1≥λ2≥…≥λ q and the corresponding eigenvectors ±e1, e2, …, e q , select the eigenvectors corresponding to the first r largest eigenvalues ​​to form the projection matrix P = [e1, e2, ..., e r ], the low frequency subband coefficient A l Projected onto the projection matrix, the reduced-dimensional feature vector IR is obtained r , the calculation formula is The feature vector IR of the infrared thermal imaging image is formed.

6. The wound image-based adaptive analysis nursing method according to claim 2 or 3, characterized in that: The fluorescence imaging image feature extraction adopts the local binary pattern LBP feature extraction algorithm to extract the texture and grayscale distribution features of the image and generate the feature vector of the fluorescence imaging image, specifically: Fluorescence imaging image I f Each pixel point p(x, y) in is taken as the center, a circular neighborhood with a radius of R is selected, and the neighborhood contains sampling points. The sampling points in the neighborhood are bilinearly interpolated to obtain the gray value gi g (i g =1, 2, ..., N), the gray value of the central pixel is g c , calculate the LBP encoding value LBP R，N (x, y), the formula is: Among them, s g (x) is a sign function, which is defined as: By traversing all the pixels in the fluorescence imaging image, the LBP feature map LBP is obtained map , obtain the texture features of the fluorescence imaging image; The grayscale distribution characteristics of the fluorescence imaging image are extracted by grayscale co-occurrence matrix; the fluorescence imaging image I f Construct a gray level co-occurrence matrix under a certain gray level G, distance d and direction θ, and the gray value in the image is i h and j h The number of times a pixel pair appears is P(i h , j h ), then the gray-level co-occurrence matrix element GLCM d，θ (i h , j h ) is calculated as: By using the gray-level co-occurrence matrix GLCM d，θ (i h , j h ) Obtain texture features; The LBP feature map LBP map Combined with the texture features extracted from the gray-level co-occurrence matrix, the feature vector F of the fluorescence imaging image is formed. Y .

7. The wound image-based adaptive analysis nursing method according to claim 3, characterized in that: The reconstructing and generating the fused image based on the fused feature vector M is specifically as follows: The fused feature vector M is input into the image reconstruction model improved by the generative adversarial network, and the image reconstruction model consists of a generator G and a discriminator D; the generator G performs image reconstruction and outputs a feature map F through the generator G. G , the size is H×W×C, where H is the height, W is the width, and C is the number of channels. The feature map F is fused through an adaptive weighted fusion layer. G The weight matrix of the adaptive weighted fusion layer is W a , according to the feature map F G The importance of features in different regions is dynamically generated. For each pixel in the feature map (i x , j x ), calculate its feature importance score S(i x , j x ), the calculation formula is where α c is the importance coefficient of channel c, and the element w of the weight matrix is ​​generated using the Softmax function according to the score a (i x , j x ), the formula is Where m represents the pixel index in the height direction of the feature map, and n represents the pixel index in the width direction of the feature map. G With the weight matrix W a Perform element-by-element multiplication and accumulation to obtain the reconstructed fused image I fusion (i x , j x ), the calculation formula is 8. The wound image-based adaptive analysis nursing method according to claim 1, characterized in that: In step S3, the fused image is analyzed using a deep learning target detection model to identify the boundary, type, area and depth information of the wound, and form a wound information set, which specifically includes: The deep learning target detection model is a model improved based on a convolutional neural network. During the model training phase, a large amount of annotated multimodal wound image data is used for training to optimize model parameters. The reconstructed fused image I fusion (i x , j x ) Input the deep learning target detection model based on Transformer improvement, and use the multi-head self-attention mechanism to capture the global features to obtain F a , which is refined by the convolutional neural network to obtain F c , through the formula P b (x j ,y j )=Sigmoid(W b F c (x j ,y j )+b b ) predicts the wound boundary, where W b is the boundary weight matrix, b b is the bias term, (x j ,y j ) is the pixel coordinate, which is greater than the threshold τ b The pixel points are boundary points; F is obtained by global average pooling p , through the fully connected layer P t (i) = Softmax(W c F p +b c ) Identify the wound type according to the number of pixels N within the wound boundary b And the pixel corresponding area S p , by A=N b ×S p Calculate the wound area, combine the multimodal image features, and use the formula Estimate wound depth, where F d is the input feature, D(x, y) is the wound depth value output, G k are different depth estimation sub-models, w k is the corresponding weight coefficient, b d is the bias term, K is the number of sub-models; The wound boundary, type, area and depth information are integrated to form a wound information set.

9. The wound image-based adaptive analysis nursing method according to claim 1, characterized in that: In S4, based on the identified wound information set, combined with the patient's personal medical history, age, and physical condition data, a personalized nursing plan is formulated through an adaptive nursing strategy generation model, specifically including: The adaptive nursing strategy generation model uses a decision tree algorithm to screen and combine the most appropriate nursing plan from the nursing strategy template based on wound information and patient personal data; The wound information set and the patient's personal data are quantified into feature vectors When building a decision tree, use Gini impurity Measures the purity of the dataset, where m is the number of categories and p is i is the proportion of samples belonging to the i-th class in the data set, according to Select the smallest G s The feature and threshold splitting nodes are as follows: D1 and D2 are two sub-datasets after the split, and |D|, |D1|, and |D2| represent the number of samples in the corresponding data sets respectively; the new feature vector After input, traverse the decision tree to the leaf node, and the node corresponds to the nursing strategy sub-template set {S1, S2, ..., S k },according to Calculate the matching score, where w i is feature x i The weight of Represents feature x i With sub-template S j The matching degree of the corresponding feature conditions in the range of [0, 1] is determined by selecting the sub-template with a high score to form a definite nursing plan.

10. The wound image-based adaptive analysis nursing method according to claim 1, characterized in that: In the step S5, wound data is collected during the nursing process and transmitted to the HIS system via Bluetooth or 5G. After data cleaning, the LSTM model is used to predict the wound healing trend. When an abnormality occurs, re-collection is triggered. Then, the data is processed and analyzed according to S12-S14 to generate a new nursing plan. According to the new plan, nursing resources are allocated using an intelligent scheduling algorithm to achieve dynamic adjustment of the nursing plan.

Images