Image segmentation method for burn wound area

Abstract

The invention relates to an image segmentation method for a burn wound area. The method comprises the following steps: collecting burn wound images under different spectra, including multiple bands of visible light, infrared light and ultraviolet light, by using a multispectral imaging device; analyzing the reflection and absorption characteristics of a burn area under different bands by using the optical characteristics of biological tissue for each spectral image, and marking the potential burn area; designing an adaptive algorithm, dynamically adjusting parameters of an imaging device including exposure and light intensity based on feedback of different spectral characteristics; enhancing the characteristic contrast of the burn area; preliminarily defining the boundary of the burn area by using a threshold segmentation or a regional growing algorithm according to the enhanced spectral image; inputting the multispectral image into a hierarchical processing model, and decomposing different tissue layers of the burn area including the epidermis and the dermis based on different bands or color information; and extracting the detail features of each layer respectively by using a feature extraction algorithm including texture analysis and edge detection on each hierarchical result.

Description

Technical Field

[0001] The invention relates to an image segmentation method, in particular to an image segmentation method for a burn wound area. Background Art

[0002] The current image segmentation method for burn wound areas, as described in Chinese invention patent 2020116447542, can provide pixel-level burn image segmentation with a certain degree of accuracy, but there are still some technical barriers in specific applications. The patented method relies on a neural network to process burn images, and improves the image quality through steps such as preprocessing and data enhancement, and finally performs segmentation processing by a model consisting of an encoder and a decoder. Although this method has achieved certain results in image processing and deep learning, there are still the following key issues in practical applications.

[0003] First, this method has deficiencies in the segmentation of burn depth and tissue layers. Burn wounds usually include multi-layer structures such as the epidermis, dermis, and subcutaneous tissue, and the depth of injury is a key indicator for the evaluation and treatment of burns. However, the segmentation method of patent number 2020116447542 relies only on two-dimensional image features and does not fully consider the changes in the longitudinal depth of the burn wound, which makes it limited in its ability to recognize deep burns. This method lacks the processing of three-dimensional depth data and cannot accurately distinguish the degree of burns at different levels, especially when accurate assessment of damage to the dermis or subcutaneous layer is required. It cannot provide sufficient information support. This may mislead the diagnosis and treatment decisions of severe burns and affect the final treatment effect.

[0004] Secondly, this method has the problem of inconsistency in the boundary processing of the segmentation results. Although the encoder and decoder architecture can extract and process image features, the segmentation results often show uneven or incoherent boundaries when processing complex boundary shapes, especially the transition area between the burn wound and the healthy skin. Burn areas are usually accompanied by irregular shapes and fuzzy boundaries. Simply relying on pixel-level image feature extraction cannot fully and accurately capture these complex boundaries. This boundary inconsistency leads to a decrease in the accuracy of the segmentation results, which may cause deviations in the design of clinical treatment plans. Especially when precise operations are required, the fuzziness of the boundaries will increase the difficulty of surgery or delay the recovery process. Third, the existing methods lack the ability to be intelligent and automated in burn grading. Although the neural network model can segment images, it mainly focuses on the morphological features of the burn area rather than the depth features. This segmentation based on two-dimensional images is difficult to effectively distinguish the depth and level of burns, and thus cannot achieve burn grading. Burn grading is crucial for developing treatment plans. The treatment methods for superficial burns and deep burns are completely different. The segmentation method of this patent can only identify the surface wound, without considering the depth of the burn extension, and lacks accurate judgment of different burn levels. This limitation requires doctors to rely on experience or additional testing methods to make judgments when developing treatment plans, which increases the uncertainty of treatment.

[0005] In addition, although the data enhancement strategy in this patented method can improve the robustness of the model and expand the diversity of training data, it does not fully incorporate the actual characteristics of burn images. In burn images, the morphology and boundary changes of the damaged area are not random, but are affected by biological tissue and the degree of damage. Simple random rotation, flipping, scaling and other operations can avoid overfitting of the model, but these operations may cause the model to be unable to effectively capture the unique features of the burn area. Especially when faced with complex burn morphologies, the data brought by the enhancement operation may deviate from the actual scene, resulting in recognition difficulties during model training. Summary of the invention

[0006] The purpose of the present invention is to provide an image segmentation method for a burn wound area, thereby solving some of the drawbacks and shortcomings pointed out in the background technology.

[0007] The present invention solves the above-mentioned technical problems by adopting the following technical solution, which includes the following steps:

[0008] S1. Adaptive spectral imaging segmentation based on optical properties of biological tissues:

[0009] S1.1. Collect burn wound images under different spectra using multispectral imaging equipment, including multiple bands of visible light, infrared light, and ultraviolet light; for each spectral image, use the optical properties of biological tissue to analyze the reflection and absorption characteristics of the burn area under different bands and mark potential burn areas;

[0010] S1.2. Design an adaptive algorithm to dynamically adjust the parameters of the imaging device including exposure and light intensity based on the feedback of different spectral features; enhance the characteristic contrast of the burn area;

[0011] S1.3. Preliminarily delineate the boundaries of the burn area based on the enhanced spectral image by threshold segmentation or region growing algorithm;

[0012] S2. Burn tissue image segmentation based on hierarchical decomposition:

[0013] S2.1. Input the multispectral image into the hierarchical processing model to decompose different tissue layers of the burn area including the epidermis and dermis based on different bands or color information;

[0014] S2.2. On each layered result, use a feature extraction algorithm including texture analysis and edge detection to extract detailed features of each layer to distinguish burn tissues at different depths;

[0015] S2.3. Based on the hierarchical features, a convolutional neural network (CNN) architecture is used to independently segment each layer of burn tissue. The segmentation results of each layer are merged to generate a complete segmentation map of the burn wound area.

[0016] S3. Boundary segmentation of the junction between deep burn wound and normal skin:

[0017] S3.1. In the segmentation results, identify the boundary area between the burn wound and the normal skin, and extract the boundary features of the area, including texture differences and optical reflection changes;

[0018] S3.2, use GraphCut or LevelSet algorithm to optimize the local boundary of the junction;

[0019] S4. Grading and segmentation of fusion burn depth and tissue structure:

[0020] S4.1. Combine optical coherence tomography (OCT) or ultrasound imaging to obtain three-dimensional depth information of the burn wound and form a preliminary depth distribution map;

[0021] S4.2. Design a deep learning model to combine the burn area segmentation results with the three-dimensional depth information to establish a burn grading model; classify the burn grades according to the depth characteristics, including epidermal burns and dermal burns;

[0022] S5. Segmentation and tracking of healing process based on temporal features:

[0023] S5.1. During the burn healing process, regularly collect wound surface images and construct time series image data of the healing process;

[0024] S5.2. Analyze the image segmentation results at consecutive time points to extract the temporal characteristics of wound area reduction, boundary changes, and healing degree;

[0025] S5.3. Dynamically optimize the segmentation model based on the time series characteristics; adjust the segmentation parameters according to the healing progress.

[0026] Furthermore, the method for generating a complete burn wound area segmentation map includes:

[0027] The burn wound is regarded as a single multi-layer structure, each layer represents a different tissue structure including the epidermis, dermis, and subcutaneous layer; each layer presents a different feature pattern in the image; the unique features of each layer are extracted to design a multi-layer convolutional neural network architecture CNN, which processes the image layer by layer and designs different convolution kernels Extract organizational features at each level; the parameters of the convolution kernel θ n The process of tuning at different levels is expressed by the following formula:

[0028]

[0029] in,

[0030] represents the feature response extracted at the nth level, (x, y) is the coordinate of the image on the two-dimensional plane; Ω represents the pixel space of the image, covering the entire image area in the integral; The convolution kernel function of the nth level, the convolution kernel parameter θ n It is an adjustable parameter corresponding to the structural features of each layer, which is used to extract the image features of this layer; I(xx′,yy′) is the pixel intensity value of the input image at the offset coordinate (xx′,yy′).

[0031] Furthermore, the method for generating a complete burn wound area segmentation map includes:

[0032] Design a feedback and correction mechanism, by sharing some feature information, take the segmentation result of the surface layer as the prior information of the next layer, and dynamically adjust the segmentation strategy of the next layer; the feedback mechanism allows the segmentation processes of different levels to influence each other, in terms of boundary processing; set the segmentation result of the previous layer as R n-1 (x, y), the deep segmentation result is corrected by extracting the gradient change boundary information; the correction formula is as follows:

[0033]

[0034] Among them, H n (R n-1 (x,y),λ n ) represents the result of the correction operation at the nth layer; R n-1 (x, y) represents the segmentation result of the n-1th level, which serves as the preliminary input of the current level; Represents the gradient of the segmentation result of the previous layer, which is used to measure the boundary change; n To control the adjustment parameters of the correction strength, it determines the influence of the upper layer segmentation results on the lower layer.

[0035] Furthermore, the method for generating a complete burn wound area segmentation map includes:

[0036] Define a global optimization objective function, and optimize based on the difference between the segmentation results of each level and the merged image; the goal of the optimization process is to eliminate boundary inconsistencies between levels by minimizing the energy function:

[0037]

[0038] Among them, J(R n ,G) is the global optimization energy function, and the multi-layer segmentation result is optimized by minimizing this function; R n (x, y) is the segmentation result of the nth level; G(x, y) is the global target segmentation result, which represents the desired complete segmented image after optimization; Ω represents the integration range, that is, the pixel space of the entire image.

[0039] Furthermore, the method for generating a complete burn wound area segmentation map includes:

[0040] Combined with three-dimensional depth data obtained by optical coherence tomography (OCT) equipment; layered processing of burn tissue in three-dimensional space involves features in the x and y dimensions, as well as changes in the depth dimension z; convolution operations are extended to three-dimensional space to process tissue features of burn wounds at different depth layers;

[0041] The three-dimensional segmentation formula is:

[0042] D θ (x,y,z)=∫ Ω k θ (x′,y′,z′)·I(xx′,yy′,zz′)dx′dy′dz′

[0043] Among them, D θ (x, y, z) is the three-dimensional feature response, which means the feature extraction result at the (x, y, z) coordinate; k θ(x′, y′, z′) is a three-dimensional convolution kernel function, θ is the parameter of the convolution kernel, and is used to extract the three-dimensional features of the image; I(x, y, z) represents the input three-dimensional burn wound image.

[0044] Furthermore, the method for establishing a burn grading model comprises:

[0045] Surface feature extraction is performed on the two-dimensional image of the burn area, including extracting local edge, color and texture information from the image to identify the burn boundary and changes in the wound morphology, as well as surface damage. The convolutional neural network calculates the response of the image through convolution operations, which is specifically expressed as:

[0046] F θ (x,y)=∫ Ω k θ (x′,y′)·I(xx′,yy′)dx′dy′

[0047] in,

[0048] F θ (x,y) represents the feature response at the coordinate (x,y), which is the feature extraction result of the convolution kernel on the input image; I(x,y) is the input two-dimensional burn image, where x and y are the horizontal and vertical coordinates of the image respectively; k θ (x′, y′) is the convolution kernel function, θ is the parameter of the convolution kernel, which is used to adaptively adjust the weight of the filter according to different levels to ensure that features of different scales can be extracted; Ω is the area of ​​image pixels, and the integral operation covers this area to extract features.

[0049] Furthermore, the method for establishing a burn grading model comprises:

[0050] Combined with the three-dimensional depth information, the vertical depth change of the burn area is analyzed through the three-dimensional convolution network; the vertical features are extracted through the three-dimensional convolution to classify the burns at different depth levels; the three-dimensional convolution operation is used to extract the depth information of the burn area, which is specifically expressed as:

[0051] D θ (x,y,z)=∫ Γ k θ (x′,y′,z′)·I(xx′,yy′,zz′)dx′dy′dz′

[0052] Among them, D θ (x, y, z) is the characteristic response at the three-dimensional coordinate (x, y, z), which is used to reflect the longitudinal depth characteristics of the burn area; z represents the depth dimension of the burn tissue, providing information about the damage to the dermis and subcutaneous tissue; I(x, y, z) is the input three-dimensional burn image, where x and y are the horizontal and vertical coordinates, and z is the depth information; kθ (x′, y′, z′) is a three-dimensional convolution kernel used to extract three-dimensional features from an image. θ is the parameter of the convolution kernel, and the convolution kernel weight is automatically adjusted according to different levels. Γ represents the three-dimensional pixel space, and the integral operation is performed in this space to capture depth information at different levels.

[0053] Furthermore, the method for establishing a burn grading model comprises:

[0054] The attention mechanism is introduced to focus on the serious part of the burn area, so as to extract and grade the features of the local area. Through the attention mechanism, the focus of the model is dynamically adjusted according to the depth characteristics of the burn area. In the global grading process, the surface and depth features are weighted and fused through the following formula to achieve grading judgment:

[0055]

[0056] Among them, G(F,D) is the global classification result, which means that based on surface features and deep features The comprehensive weighted result of It is the surface feature response of the two-dimensional image at the i-th level, reflecting the information of the burn surface; is the depth feature response of the three-dimensional image at the i-th level, reflecting the longitudinal information of the burn depth; w i is the feature weight coefficient, which is used to adjust the importance of surface features and depth features; w i The value of is adjusted through model learning; N is the total number of model levels, indicating the number of fusions at different feature levels.

[0057] The burn wound area image segmentation method of the present invention has the following beneficial effects:

[0058] Through a multi-level convolutional neural network architecture, combined with two-dimensional and three-dimensional image data, different layers of the burn area are segmented. This method can accurately capture the boundary characteristics and depth changes of the epidermis, dermis and subcutaneous tissue, avoiding the problem of inconsistent boundaries between layers that is prone to occur in traditional methods, and significantly improving the accuracy of segmentation. Through a deep learning model, combined with the surface features of two-dimensional images and the depth features of three-dimensional images, different injury layers are automatically identified, and the burn grade is intelligently graded, which reduces the subjective errors of doctors and improves the efficiency and accuracy of burn grading.

[0059] The feedback and correction mechanism is adopted to dynamically adjust the segmentation strategy of the next layer by taking the segmentation results of the surface layer as the prior input of the deep layer. This mechanism ensures the mutual influence between the segmentations of different levels, especially improving the consistency and coherence of the segmentation at complex boundaries. By minimizing the energy function, the global optimization of the segmentation results of each level is achieved, the boundary discontinuity between different levels is eliminated, and the generated final segmentation image is ensured to have smooth and coherent boundaries, providing more accurate burn wound images for subsequent treatment decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 The figure is a flow chart of the image segmentation method of the burn wound area of ​​the present invention.

[0061] Figure 2 The present invention provides a flow chart of the method for generating a complete burn wound area segmentation map.

[0062] Figure 3 The present invention provides a flow chart of the method for establishing a burn grading model.

[0063] Figure 4 It is a real picture of foot burns of the present invention. DETAILED DESCRIPTION

[0064] The specific implementation modes of the present invention will be described in detail below in conjunction with the accompanying drawings.

[0065] Combined with Figure 1 As shown in the flowchart, the core of the image segmentation method of the burn wound area lies in the adaptive spectral imaging segmentation based on the optical properties of biological tissue. The specific steps are: S1. Collect burn wound images under different spectra through multi-spectral imaging equipment, and the collected spectral range includes multiple bands such as visible light, infrared light and ultraviolet light. Under each spectrum, the optical properties of biological tissue, such as reflectivity and absorption characteristics, will be significantly different. Using this, the reflection and absorption characteristics of the burn area in each band can be analyzed, especially the differences between damaged tissue and healthy tissue under different spectra. Through this analysis, the system can mark potential burn areas and identify injuries of different depths and degrees.

[0066] Next, by designing an adaptive algorithm, the system will dynamically adjust the parameters of the imaging device based on the feedback results of different spectral features. Specifically, key parameters such as exposure and light intensity will be automatically adjusted based on the feedback of each spectral image to ensure that the imaging device can capture the clearest and highest contrast images. The core purpose of this adaptive adjustment is to enhance the characteristic contrast between the burn area and the surrounding healthy tissue, so that the damaged area can be distinguished more accurately in the subsequent segmentation process.

[0067] After acquiring the enhanced spectral image, the system will process the image using threshold segmentation or region growing algorithm. Both algorithms are basic methods for image boundary recognition, where threshold segmentation distinguishes the burn area from the healthy area by setting a grayscale threshold, while the region growing algorithm gradually covers the entire range of the burn area by expanding from the seed point.

[0068] Step S2 is based on the burn tissue image segmentation technology of hierarchical decomposition. First, the multispectral image obtained by the multispectral imaging device is input into the hierarchical processing model. The core of this model is to decompose and process the different tissue layers of the burn area based on multi-band or different color information. Specifically, by analyzing the tissue characteristics reflected in the multispectral image, the burn area can be divided into layers of different depths, mainly including different tissue structures such as the epidermis and dermis. Through this hierarchical processing, the differences in depth and optical properties of different tissues can be captured more accurately, thus laying the foundation for subsequent fine segmentation.

[0069] After the layering process is completed, specific feature extraction algorithms are used to analyze and process the image results at each level. These algorithms include texture analysis and edge detection, which can extract detailed features of the burn area from the image. For example, through texture analysis, the surface features of the burn tissue, such as roughness, morphological changes, etc., can be identified; through edge detection, the boundary between the tissue layer and the surrounding healthy tissue can be accurately identified. These detailed features can help distinguish burn tissues of different depths at each level, thereby ensuring the accuracy of segmentation.

[0070] After obtaining the layered processing results and detailed features, the convolutional neural network (CNN) architecture is then used to independently segment each layer of burn tissue. Through multi-layer convolution operations, the CNN architecture can automatically learn and extract the spatial and morphological features of burn tissue, and segment the burn area based on these features. For each layer, the CNN model will perform independent segmentation operations based on the characteristics of the layer, ensuring that the segmentation of each layer can reflect the actual boundaries and damage of the tissue layer.

[0071] Step S3 involves the boundary segmentation of the junction between the deep burn wound and the normal skin. In the preliminary results of the segmentation, it is necessary to focus on identifying the junction area between the burn wound and the normal skin. This junction area usually has significant feature differences, mainly manifested in changes in texture and optical reflectance properties. Therefore, in this step, the boundary features of the junction area must be extracted first. These features include texture differences. Specifically, burn tissue is often rougher in morphology, while the texture of normal skin is relatively smooth. Optical reflectance properties are also an important indicator for differentiation. The reflectance and absorption rate of the burn area under different light are different from those of healthy skin, especially in the visible light and infrared light spectra. This difference is more obvious.

[0072] After completing the extraction of boundary area features, it is necessary to optimize the local boundaries of these junctions. At this time, the GraphCut algorithm or the LevelSet algorithm can be used to perform fine processing of the boundaries. The GraphCut algorithm is a graph-based segmentation method that represents the pixels of the image as nodes in the graph and uses the idea of ​​energy minimization to segment the burn area and the normal area. By setting the boundary conditions and the correlation between the image pixels, the GraphCut algorithm can accurately identify the boundaries of the junction area and optimize the boundaries to make them more consistent with the actual burn morphology.

[0073] Another commonly used method is the Level Set algorithm, which represents the segmentation boundary as an implicit function and automatically adjusts the shape of the boundary as it evolves. This method can dynamically handle the complex boundary between the burn area and normal skin, and is particularly suitable for processing areas with irregular shapes and blurred boundaries. Through the Level Set algorithm, the system can continuously adjust the position of the boundary according to the local gradient and boundary features of the image during the segmentation process to ensure that the segmentation boundary of the junction area is accurate and smooth.

[0074] Step S4 involves hierarchical segmentation that integrates burn depth and tissue structure. In this process, three-dimensional depth information of the burn wound is first obtained through non-invasive techniques such as optical coherence tomography (OCT) or ultrasound imaging. These imaging technologies can provide accurate depth data, especially in terms of resolution of different tissue layers, which can reflect the deep damage in the burn area. Through these technologies, the system can generate a preliminary depth distribution map of the burn wound, which can clearly show the different depth characteristics of the burn from the epidermis to the dermis. This step is the basis of the entire hierarchical segmentation because it can provide structural information of the burn wound at different depths and help distinguish between superficial and deep injuries.

[0075] After obtaining the depth distribution map, the next step is to design a deep learning model to fuse the segmentation results of the burn area with the acquired three-dimensional depth information. The core function of the deep learning model is to automatically learn the characteristics of the burn area from the data, and accurately divide the damage at different levels in combination with the three-dimensional depth information. The model processes multidimensional data through convolutional neural networks (CNN) or three-dimensional convolutional neural networks (3DCNN), automatically extracts the characteristics of the burn area, and fuses the depth information at the same time, so that the segmentation results are not limited to the surface, but can also accurately reflect the damage to the deep tissue. By simultaneously processing the image data and depth data of the burn area, the model can establish a complete burn grading framework.

[0076] In this grading framework, the system will automatically classify the burn grade according to the depth characteristics of the burn area. The classification of burn grades is mainly based on the degree of damage to different tissue layers. For example, epidermal burns are usually limited to the outermost layer of the skin, while dermal burns affect deeper tissue layers and may involve damage to nerve endings and blood vessels. The deep learning model can automatically identify these different layers based on three-dimensional depth data and generate corresponding burn grades. This grading method based on deep features is not only more accurate, but also can reduce the subjective errors caused by human judgment, ensuring the scientificity and consistency of burn grading.

[0077] Step S5 is based on the segmentation and tracking of the healing process based on time series features, which is mainly used to continuously monitor and track the dynamic changes of the wound area during the burn healing process. This step first requires the regular collection of image data of the burn wound at different healing stages to form a complete time series image data set. These images can capture the changes of the wound at different time points during the healing process, providing basic data for subsequent analysis and model optimization. The acquisition of time series data is usually based on fixed time intervals, such as regularly recording the changes in the wound every day or every week to ensure that the entire healing process can be fully reflected.

[0078] Next, the system will analyze the image segmentation results at each time point and extract the temporal features of the healing process. Specifically, the reduction of wound area and the change of boundaries are two key healing indicators. During the healing process, the area of ​​the burn wound area usually decreases gradually, and the shape and clarity of its boundaries also change. By tracking and analyzing these features, the system can evaluate the degree of wound healing, including the speed of wound reduction, the smoothness of the boundaries, and morphological changes. The segmentation results at each time point not only reflect the wound state at that time, but also provide the trend of changes between consecutive time points, thereby providing dynamic temporal features for the healing process.

[0079] Based on the extracted time series features, the system will dynamically optimize the segmentation model. As the healing process progresses, the wound morphology will continue to change, and the initial segmentation model may no longer be applicable to the later healing stage. Therefore, the segmentation model needs to be dynamically adjusted according to the features in the time series to adapt to the gradual changes in wound morphology. The system adjusts the model's segmentation strategy by analyzing the feature differences at different time points to ensure that the segmentation results can accurately reflect the healing status of each stage.

[0080] Finally, the system automatically adjusts the segmentation parameters based on the progress of healing. Specifically, during the healing process, the boundaries of the wound surface will gradually smooth out and the area will shrink, which means that the parameters of the segmentation model also need to be adjusted accordingly to adapt to these changes. For example, the segmentation model may adjust the sensitivity of boundary detection to ensure that the area can still be accurately divided when the wound surface shrinks.

[0081] Embodiment 1:

[0082] Combined with Figure 2 As shown in the flowchart, a large hospital is treating a severely burned patient. The patient's back, arms and other parts have suffered third-degree burns, and the wound surface involves multiple layers of tissue, including the epidermis, dermis and subcutaneous tissue. During the treatment process, the doctor needs to accurately assess the depth of the burn in order to formulate a subsequent treatment plan. The image segmentation method of the present invention can help doctors accurately segment the burn area and assess the damage at different levels, thereby providing scientific data support for doctors.

[0083] In this case, the hospital used a set of multispectral imaging equipment to collect images of the patient's burn wound. The equipment can capture multi-level structural information of the burn area, including the epidermis, dermis, and subcutaneous tissue. Through multispectral imaging, burn images in different bands were obtained, which clearly showed the differences in optical characteristics of different layers of tissue. The epidermis showed obvious pigmentation and damaged areas in the visible light band, the dermis reflected the damage to blood circulation in the infrared band, and the changes in the subcutaneous tissue could be observed through ultrasound images or infrared deep imaging technology.

[0084] In order to achieve accurate segmentation of the burn wound area, the system designed a multi-level convolutional neural network (CNN) architecture. This architecture is specifically designed to process image data of multi-level tissues. First, the input multispectral image is processed in layers to extract the unique features of each layer. Specifically, the feature extraction of the epidermis requires a specific convolution kernel to capture the characteristics of the epidermis such as pigment changes and damage morphology; the dermis uses another set of convolution kernels to capture deep features such as impaired blood flow and tissue penetration. In order to accomplish these tasks, the design of the convolution kernel and the tuning of the parameters are crucial.

[0085] In this case, the convolution kernel parameters θ are set n To match the characteristics of each layer of tissue. For the epidermis, a smaller convolution kernel size, θ1=3×3, is selected to process small surface damage features; and for the dermis, a larger convolution kernel size, θ2=5×5, is selected to better capture large areas of damaged blood flow in the dermis. Similarly, in the processing of the subcutaneous layer, a more complex convolution kernel, θ_3=7×7, can be selected to deal with deeper tissue changes. In the actual calculation process, the feature extraction of each level is expressed by the following formula:

[0086]

[0087] in, Represents the feature response extracted from the nth layer. For example, when n=1, it represents the epidermis, when n=2, it represents the dermis, and when n=3, it represents the subcutaneous layer. is a function of the convolution kernel, Ω represents the pixel space of the image, and covers the entire image area in the integral. The pixel intensity value I(xx′,yy′) of the image represents the pixel value at the offset coordinate (xx′,yy′). The size of the convolution kernel determines the range of features extracted by the network at each level.

[0088] Let's assume that during the segmentation process, the system detects that the average pixel intensity of the epidermis ranges from 150 to 200, while the average pixel intensity of the dermis ranges from 100 to 150. For the subcutaneous layer, due to the effect of light penetration, the pixel intensity fluctuates between 50 and 100. By using different convolution kernel sizes and adaptive tuning of the deep learning model, the system can accurately identify the boundaries of different layers, especially the transition area between burns and healthy tissues.

[0089] In the final generated segmentation map, the burn area of ​​the epidermis occupies 30% of the image, the burn area of ​​the dermis occupies 50% of the image, and the damage to the subcutaneous layer is relatively small, occupying 20% ​​of the image. This segmentation result can provide doctors with a complete distribution of the burn area and accurately assess the depth of the injury.

[0090] In actual operation, the tuning range of the convolution kernel parameters can be adjusted according to the specific tissue level. For the epidermis, the value range of θ1 can be between 2×2 and 4×4, while for the dermis, θ2 can be between 4×4 and 6×6. The convolution kernel parameter θ3 of the subcutaneous layer is usually larger, and the value range can be between 6×6 and 8×8. The adjustment of these parameters depends on the complexity of each layer of tissue and the required resolution. Through these steps and calculations, the final generated complete burn wound area segmentation map can accurately reflect the damage of each layer of tissue.

[0091] In order to further improve the segmentation accuracy, the system introduces a feedback and correction mechanism, especially for the processing of the complex boundary between the burn area and the healthy tissue. In the segmentation process, the role of feedback and correction mechanism is particularly important. Since the characteristics of burn wounds at different levels are obviously different, for example, the damage morphology of the epidermis is more superficial, while the dermis involves deeper damage, and the boundaries are fuzzy and complex. Therefore, the system has designed a dynamic adjustment strategy that allows the layers to influence each other during the segmentation process. Specifically, the system uses the segmentation results of the epidermis as the prior input of the dermis, and uses the gradient changes of the surface layer to correct the segmentation results of the deep tissue.

[0092] Assume that the current processing is the segmentation of the dermis layer, and the segmentation result of the previous epidermis layer is R n-1 (x, y), the segmentation result has clearly defined the boundary of the epidermal damage area. In order to correct the segmentation result of the dermis, the system extracts the gradient change of the epidermal boundary This step is crucial because gradient changes can be used to capture the boundary area between the epidermis and healthy skin. In the boundary area, the gradient changes fluctuate between 0 and 10, where larger gradient values ​​represent significant boundary changes. For deeper tissues, the system needs to guide the segmentation strategy based on this gradient information to ensure that the segmentation results of the dermis do not deviate from the actual tissue boundary.

[0093] The specific correction process is expressed by the following formula:

[0094]

[0095] Here, H n (R n-1 (x,y),λ n ) represents the correction operation result of the current layer (i.e., the dermis layer), which is dynamically adjusted according to the gradient changes of the previous layer. n-1 (x, y) represents the segmentation result of the previous layer, that is, the segmentation boundary of the epidermis. In order to ensure that the correction operation has sufficient impact on the current layer, the system also introduces an adjustment parameter λ n The function of this parameter is to control the influence of the previous layer segmentation result on the current layer. Usually, λ n The value range of is between 0.1 and 1.0. The larger the value, the stronger the influence of the previous layer segmentation on the current layer. In this case, in order to ensure that the dermis segmentation can closely follow the boundary of the epidermis, the system sets λ n = 0.7, which means that the segmentation result of the epidermis will have a strong influence on the dermis, but will not completely dominate.

[0096] Next, by changing the gradient With the adjustment parameter λn Multiplying, the system generates the correction result H of the current level n , and applies it to the segmentation of the dermis. Through this feedback mechanism, the segmentation boundary of the dermis is not only segmented independently according to its own characteristics, but also corrected by the segmentation information of the previous layer. For example, if the gradient change of the epidermis is large in a certain local area, the segmentation boundary of the dermis will move closer to this area, thus ensuring the consistency of the segmentation results between different layers.

[0097] To verify the effectiveness of this correction strategy, doctors compared the segmented dermis region with actual clinical images. The results showed that after the introduction of the feedback mechanism, the boundary between the dermis and the epidermis became smoother, reducing cross-layer segmentation errors. For example, the segmentation boundary of the epidermis had obvious bumps in some areas, which were effectively corrected in the dermis segmentation through gradient information, avoiding excessive irregularity of the boundary.

[0098] Since the tissue characteristics and boundaries of each layer are different, discontinuity or overlap between different layers may occur. Therefore, the present invention introduces a global optimization step to ensure smooth transition and consistency of the segmented image at all levels of boundaries by minimizing the energy function.

[0099] In this process, a global optimization objective function is first defined. The purpose of this function is to measure the segmentation results R at different levels. n (x, y) and the final desired global target segmentation result G(x, y). To achieve this goal, the optimization process uses an energy function that optimizes the result by minimizing the error between different levels of segmentation and the global segmentation image. The mathematical representation of the process is as follows:

[0100]

[0101] Among them, J(R n ,G) is a globally optimized energy function, which represents the overall difference between the segmentation result and the target image. n (x, y) represents the segmentation result of the nth level, G(x, y) is the final desired global segmentation image, which represents the complete segmentation map after optimization of all levels. The integration range Ω represents the pixel space of the entire image, and the integration operation will be performed on all pixels to ensure that the optimization result covers the entire image.

[0102] In this case, the patient's burn wound is divided into three layers: epidermis, dermis, and subcutaneous layer. The previous steps have generated the segmentation results of these three layers, namely R1(x, y), R2(x, y), and R3(x, y). These segmentation results may be inconsistent at the boundaries of some areas. For example, the boundary of the epidermis may not completely overlap with the boundary of the dermis, resulting in overlap or gaps in the segmented images. To solve this problem, the system minimizes the above energy function J(R n ,G) to optimize the segmentation results of each level, making the boundaries of all levels smoother and eliminating the differences between levels.

[0103] In practical applications, the pixel values ​​of the epidermis segmentation R1(x, y) in the boundary area are set to fluctuate between 200 and 255, the pixel values ​​of the dermis segmentation R2(x, y) are set to fluctuate between 150 and 200, and the pixel values ​​of the subcutaneous segmentation result R3(x, y) are set to fluctuate between 100 and 150. According to the doctor's expectations, the final global segmentation result G(x, y) should be consistent in the boundary area, for example, the transition of pixel values ​​between the dermis and the epidermis should be smooth. The system minimizes J(R n ,G) to adjust these segmentation results to ensure that in the final G(x,y), the boundary value gradually transitions from the epidermis to the dermis without any mutation or discontinuity.

[0104] To further specify, an initial energy function value is set in the optimization process. For example, in the initial stage of segmentation, the boundary difference between the epidermis and the dermis is large, and the energy function value J may be 5000. Through the optimization process, the system gradually reduces the boundary error of each layer, and finally the energy function value after optimization drops below 500, indicating that the boundary difference between different layers has been significantly reduced.

[0105] During the optimization process, the system dynamically adjusts the global target segmentation image G(x,y) based on the segmentation results of each layer. For example, if the boundary of the dermis layer is set to have a large deviation in a certain area (for example, the pixel value is 180, while the global target is 200), the system will adjust the segmentation result of this area to make it closer to the global target by minimizing the energy function. At the same time, similar adjustments will be made on the boundaries of other layers to ensure that the segmentation results of all layers can be seamlessly integrated.

[0106] In this process, the parameter λ is adjusted n It also plays a role in the aforementioned feedback and correction mechanism. During the optimization process, the system adjusts λ according to the boundary differences at each level. nFor the epidermis, the parameter λ1 may be set to a range of 0.5 to 0.8, because the characteristics of the epidermis change more obviously and require a greater correction; for the dermis and subcutaneous layer, the parameters λ2 and λ3 can be set to a range of 0.3 to 0.7, because the changes in these layers are relatively mild, but a certain degree of boundary correction is still required. Ultimately, through this global optimization method, the doctors obtained a complete segmentation map of the burn wound area with smooth and continuous boundaries.

[0107] In order to more accurately assess the damage in the deep burn area, it was decided to further combine the three-dimensional depth data obtained by the optical coherence tomography (OCT) device to achieve more accurate layered processing.

[0108] OCT devices provide detailed 3D depth data, and by acquiring information in the z dimension, doctors can further assess the depth of burns. This is particularly important for identifying injuries deep in the dermis or subcutaneous layer, as these areas are difficult to accurately segment in 2D images. Using 3D data, lateral information in the x and y dimensions, as well as depth information in the z dimension, can be processed simultaneously to ensure that all layers are accurately distinguished.

[0109] To achieve this goal, the system uses a convolution operation that is extended to three dimensions. The depth dimension z is added to the two-dimensional image. The system processes the input three-dimensional burn image through a three-dimensional convolution kernel to extract the tissue features of each layer. The mathematical expression of this process is:

[0110] D θ (x,y,z)=∫ Ω k θ (x′,y′,z′)·I(xx′,yy′,zz′)dx′dy′dz′

[0111] In this formula, D θ (x, y, z) is the three-dimensional feature response, which represents the features extracted from the input image at the (x, y, z) coordinates; k θ (x′, y′, z′) is a three-dimensional convolution kernel, which is used to extract tissue features at different depth levels. The parameters θ of the convolution kernel control its size and range of action, and are adjusted according to the needs of different levels. I(x, y, z) represents the input three-dimensional burn wound image, where x and y are the coordinates of the two-dimensional plane and z represents the depth information. This convolution operation can process the three-dimensional spatial features of the image and help identify deep-level injuries.

[0112] Let's analyze it further through specific data. In actual operation, the OCT device provides depth information for each pixel. Set in a local area, the depth range of the epidermis is z = 0 to z = 0.1 mm, the depth range of the dermis is z = 0.1 to z = 1 mm, and the depth range of the subcutaneous layer is z = 1 to z = 2 mm. The input image I (x, y, z) contains this depth information, and the three-dimensional convolution kernel k θ The parameters θ of (x′, y′, z′) will be adjusted according to different layers. The convolution kernel θ1=3×3×3 is set for the epidermis, θ2=5×5×5 for the dermis, and θ3=7×7×7 is selected for the subcutaneous layer because deeper tissue needs to be processed.

[0113] Through the three-dimensional convolution operation in the formula, the system extracts features at different depths. When z=0.05mm in the epidermis, the extracted feature response value is 220, indicating that the damage in this area is relatively superficial; while at z=0.5mm in the dermis, the feature response value is 180, indicating that the damage in the dermis is deeper. For the subcutaneous layer, at z=1.5mm, the feature response value is 130, indicating that the deep tissue has been damaged to a certain extent.

[0114] When further segmenting, the system uses these three-dimensional feature information to accurately delineate the damaged area at each level. For example, in the epidermis, the feature boundaries extracted by the system are clear and can accurately segment the damaged epidermis area; while in the dermis, because the feature response value shows that the damage is deeper, the system will generate a larger segmentation area.

[0115] In order to ensure the consistency of the segmentation boundaries of each layer, the system also combines the above-mentioned global optimization method. Assuming that there is a certain inconsistency between the boundaries of the dermis and the subcutaneous layer in the preliminary segmentation results (for example, in some areas, the depth division of the dermis does not completely match that of the subcutaneous layer), the system minimizes the global energy function J(R n ,G) Further optimize these discontinuous boundaries to make the segmentation results smoother.

[0116] After the final segmented image is generated, doctors can clearly see the patient's injury distribution based on the three-dimensional depth information, and the superficial damage of the epidermis, the deep damage of the dermis, and the degree of damage to the subcutaneous layer are all clearly displayed. These segmentation results provide doctors with more comprehensive burn assessment data, helping them make more accurate decisions during treatment.

[0117] For the range of convolution kernel parameters, the convolution kernel parameter θ1 of the epidermis can be set between 2×2×2 and 4×4×4, the convolution kernel parameter θ2 of the dermis can be set between 4×4×4 and 6×6×6, and the convolution kernel θ3 of the subcutaneous layer is usually larger, set between 6×6×6 and 8×8×8. By adjusting these parameters, the system can adaptively adjust the 3D segmentation process to ensure that the damaged areas at different levels can be accurately divided and analyzed.

[0118] Embodiment 2:

[0119] Combined with Figure 3 In the flowchart shown, a patient has a second-degree burn on his arm, and the doctor needs to grade and assess the extent of the wound to develop a further treatment plan. To achieve this goal, the system first extracts surface features from the two-dimensional image of the burn area to analyze the boundaries, color, and texture changes of the wound. This information is crucial for determining the severity of the surface damage.

[0120] In this process, the system uses convolutional neural networks (CNN) to extract local features of the image. Through the convolution operation, the system can capture the edge information, color changes and texture details of the burn area, thereby identifying the boundary shape of the wound, the depth of the injury, etc. The patient's burn image size is set to 512×512 pixels, and the system needs to run a convolution operation on the image to extract relevant features.

[0121] The mathematical representation of the convolution operation is as follows:

[0122] F θ (x,y)=∫ Ω k θ (x′,y′)·I(xx′,yy′)dx′dy′

[0123] Among them: F θ (x, y) represents the feature response at the coordinate (x, y), that is, the feature value obtained after the convolution kernel processing. This feature response can reflect the boundary information or texture change of the burn area at that point.

[0124] I(x,y) is the input two-dimensional burn image, where x and y are the horizontal and vertical coordinates. For the burn image in this case, I(x,y) contains the pixel intensity values ​​of the wound area, which represent the color and texture changes of the burn area.

[0125] k θ(x′, y′) is the convolution kernel function, and the parameter θ controls the size and shape of the convolution kernel, which is used to adjust the weight of the filter according to different feature scales. The convolution kernel can be designed differently according to the specific level. For example, a 3×3 convolution kernel can be used for edge detection, while a 5×5 convolution kernel can be used for extracting texture details. Ω represents the pixel area of ​​the image, and the integral operation covers the entire image in order to extract features from it.

[0126] In this case, the system needs to identify the boundary of the burn wound to determine the boundary between the burn area and the healthy skin. Through the convolution operation of the convolution kernel, the system can identify significant changes in pixel values ​​on the boundary. For example, the system can detect that the color of the burn area gradually transitions from healthy light pink to dark red after the burn. This color difference will be reflected in the feature response F through the convolution operation. θ It is expressed in (x,y).

[0127] In the boundary area, the feature response value extracted after the convolution kernel operation is 180, indicating that the color and texture characteristics of the area have changed significantly, which may indicate that this is the junction of the burn and healthy skin. In the center of the wound, the feature response value is 220, indicating that the damage is more serious and the texture changes are larger. For the surrounding healthy skin area, the feature response value is less than 100, indicating that these areas have no obvious burn characteristics.

[0128] To further refine the results, the system can extract features at different levels based on different convolution kernel sizes and parameters θ. For example, when it is necessary to capture the edge of a wound, the system will select a smaller convolution kernel, such as θ = 3 × 3, to ensure that the details at the boundary are accurately captured. When detecting texture changes, the system can select a larger convolution kernel θ = 5 × 5 to identify the surface roughness of the wound or scar formation.

[0129] In the actual operation process, the system generates a feature map after each convolution operation. The size of the burn area is set to 100×100 pixels. After the convolution operation, the system generates a feature map showing that the pixel values ​​with edge changes in the area are 180 to 220, while the pixel values ​​of the healthy area remain between 50 and 100. Through this feature extraction, the system can accurately segment the boundaries of the burn area and provide data support for the subsequent burn grading model.

[0130] Next, the system will use these feature data to grade the burn wound. Areas with feature response values ​​exceeding 200 are classified as severe injuries, while areas with feature response values ​​between 150 and 200 are defined as moderate injuries. The system divides the entire burn area into different grades based on these feature response values, and ultimately generates a burn grading map to provide doctors with an accurate basis for wound assessment.

[0131] In the setting of convolution kernel parameters, for small-size edge detection, the size of the convolution kernel θ can be set between 2×2 and 4×4, while for texture analysis of larger areas, the size of the convolution kernel θ is usually set between 4×4 and 6×6. By adjusting different parameters, the system can adaptively extract features at all levels and provide comprehensive feature data for burn grading. Ultimately, doctors can clearly see the boundaries and degree of damage of the burn area based on the graded images generated by the system. This image can provide a clear direction for subsequent treatment, such as specific surgery or rehabilitation treatment for severely burned areas, and can also track the progress of healing through images.

[0132] After analyzing the two-dimensional images of the patient's arm burn area and extracting surface features, the doctor found that the depth of the burn needed to be further evaluated, especially the degree of damage to the dermis and subcutaneous tissue. At this time, it was decided to combine the three-dimensional depth information and analyze the longitudinal depth characteristics of the burn area through a three-dimensional convolutional network. This step is particularly important because the depth of the burn directly determines the treatment plan. Superficial burns can usually be recovered through local care, while deep burns may require more complex treatments such as surgery.

[0133] First, the doctor uses an optical coherence tomography (OCT) device to obtain a three-dimensional depth image of the burn area on the patient's arm. The image contains horizontal and vertical position information represented by x and y coordinates, as well as depth information represented by z coordinates. Through this system, the image not only shows the surface damage of the burn area, but also reflects the complete depth changes from the epidermis, dermis to subcutaneous tissue. The depth of the burn goes from the surface to several millimeters under the skin, and deep tissue changes are crucial for doctors to judge the degree of the burn.

[0134] In order to extract these three-dimensional depth features, the system processes the image through a three-dimensional convolutional network (3DCNN) to extract the longitudinal depth change features. The three-dimensional convolution operation uses the convolution kernel k θ (x′, y′, z′) is convolved with the input three-dimensional burn image I(x, y, z), and the result is the three-dimensional feature response D θ (x, y, z), this characteristic response reflects the depth characteristics of the burn area. The specific mathematical expression is as follows:

[0135] D θ (x,y,z)=∫ Γ k θ (x′,y′,z′)·I(xx′,yy′,zz′)dx′dy′dz′

[0136] In this formula:

[0137] D θ(x, y, z) is the characteristic response at the three-dimensional coordinate (x, y, z), reflecting the longitudinal depth characteristics of the burn tissue. By extracting features of tissues at different depths, the system can classify different burn grades.

[0138] I(x,y,z) is the input three-dimensional burn image, where x and y are the horizontal and vertical coordinates, and z represents the depth information, reflecting the tissue changes from the epidermis to the dermis. θ (x′, y′, z′) is a three-dimensional convolution kernel, and the parameter θ is used to automatically adjust the size and weight of the convolution kernel to adapt to feature extraction at different depths. Γ represents the three-dimensional pixel space, and the integration operation covers the entire image pixel space to ensure that features at different depths can be captured.

[0139] To further illustrate the feasibility of the process, the 3D image size of the burn area of ​​the patient's arm was set to 256×256×10 pixels (where 256 represents the resolution in the horizontal and vertical dimensions, and 10 represents the number of slices in the depth direction). In the convolution operation, the system first performs convolution on the epidermis, using a smaller convolution kernel. To extract the subtle damage features of the epidermis. For example, the depth range of the epidermis is between z = 0 and z = 0.1 mm. After the convolution operation, the feature response value extracted is 220, indicating that the damage in this area is obvious.

[0140] Next, the system uses a larger convolution kernel for the dermis Extract deep features. The depth range of the dermis is z = 0.1 to z = 1 mm. At this level, the extracted feature response value is 180, indicating that there is also some damage in this area, but the degree of damage is less than that of the epidermis. Finally, the system processes the subcutaneous layer and uses a larger convolution kernel k θ3 =7×7×7, extracting the features of deep tissue. The depth range of the subcutaneous layer is z=1 to z=2 mm, and the feature response value is 120, indicating that the damage in this area is relatively mild.

[0141] By extracting longitudinal depth features through three-dimensional convolution, the system can classify burn areas at different levels and combine their depth features with the burn grading model. In this process, the parameters θ of the convolution kernel are automatically adjusted according to the characteristics of each level. For example, for the epidermis, the value range of the parameter θ1 of the convolution kernel can be set to 2×2×2 to 4×4×4 to accurately capture surface details. For the dermis and subcutaneous layer, the range of the values ​​of the convolution kernel sizes θ2 and θ3 is larger, which can be set to 4×4×4 to 6×6×6 and 6×6×6 to 8×8×8, respectively, in order to extract changes in deep tissues.

[0142] In actual applications, doctors can clearly see the damage of the burn area at different depths by viewing the segmentation results generated by the system. For example, the damage to the epidermis is serious and has clear boundaries, while the damage to the dermis is relatively light, and the subcutaneous tissue is almost unaffected. Through these depth features, the system automatically divides the burn area into different levels to help doctors further clarify the treatment plan.

[0143] The final segmentation result shows that the severe burn area of ​​the epidermis accounts for 30% of the entire burn area, the moderate damage of the dermis accounts for 50%, and the mild damage of the subcutaneous layer accounts for only 20%. This segmentation result provides doctors with specific burn grading data and a clear decision-making basis for subsequent treatment. Based on these grading data, doctors can decide to perform different treatments for burn areas of different depths, such as local repair of the epidermis and deeper rehabilitation treatment for the dermis.

[0144] In order to more accurately identify the severely burned areas, doctors decided to introduce an attention mechanism. This mechanism can help the system automatically focus on the most severely damaged parts of the burn area, especially in some local areas. Through the attention mechanism, the system can identify areas with different burn depths and focus more on those severely injured areas that need priority treatment.

[0145] In this case, the patient's arm had a large burn area, involving the epidermis, dermis, and part of the subcutaneous tissue. Although the entire burn area requires attention, the severity of the burn is not uniform. For example, some areas have more severe damage to the epidermis, while other areas may only involve minor subcutaneous tissue damage. Therefore, through the attention mechanism, the system can dynamically adjust the focus of the model according to the depth features and focus on the most serious areas. Through in-depth analysis of these local areas, doctors can accurately assess the depth and extent of the injury and make graded judgments based on the results of this analysis.

[0146] To achieve this function, the system uses a weighted fusion model to combine surface features With deep features Combined, the global grading result G(F,D) is generated through weighted fusion. The formula is as follows:

[0147]

[0148] In this formula:

[0149] G(F,D) represents the global classification result, which is the weighted comprehensive result of surface features and depth features. Through this classification result, the system can determine the burn grade of each area. It is the surface feature extracted from the two-dimensional image at the i-th level, reflecting information such as color and texture changes on the burn surface. It is the depth feature extracted from the 3D image at the i-th level, reflecting the longitudinal depth change of the burn, especially the degree of damage from the epidermis to the subcutaneous tissue.

[0150] w i is the feature weight coefficient, which is used to adjust the weight between surface features and depth features. For areas with heavy surface damage, the system will automatically increase w i For areas with more serious deep damage, w i These weights are adjusted dynamically through the learning of the model. N is the total number of layers in the model, indicating the number of times the system fuses at different feature levels. Assuming that the system is divided into three layers: epidermis, dermis, and subcutaneous layer, then N = 3.

[0151] Let's use actual data to further illustrate. In the 3D image of the burn area on the patient's arm, the damaged area of ​​the epidermis is extracted through surface feature extraction, and the feature response The value range of is 200 to 240, indicating that the epidermis is severely damaged. For the dermis, the depth feature The values ​​range from 160 to 200, reflecting moderate damage to the dermis. The value is only 120 to 150, indicating that the damage to the subcutaneous tissue is relatively minor.

[0152] When performing weighted fusion, the system automatically adjusts the weight coefficient w according to the degree of damage. i For the epidermis, the system sets w1 = 0.7, which means that the damage to the epidermis is more serious, and the system will focus more on the surface features. For the dermis, the system sets w2 = 0.5 to balance the influence of surface and depth features. For the subcutaneous layer, due to the less damage, the system sets w3 = 0.3, which means that the influence on the depth features is less during grading.

[0153] Through calculation, weighted fusion generates a global grading result. For a severely burned area, the system's global grading result G(F,D) may be 230, indicating that the burn level in this area is high; while for another less severely burned area, the grading result G(F,D) is only 150, reflecting that the damage in this area is less severe. These grading results help doctors accurately classify each burn area, thereby determining the areas that need priority treatment and formulating appropriate treatment plans.

[0154] To further verify the feasibility of this grading model, doctors can monitor the changes in the patient's healing process in real time. On the third day after treatment, the damage to the epidermis gradually subsided, and the surface characteristics The value of drops to 180, while the deep features of the dermis remain at around 160. By adjusting the weights w1 and w2, the system can automatically update the global grading results to reflect the healing progress of the burn. This process ensures that the system can dynamically adapt at different healing stages and provide doctors with real-time burn assessments. In summary, the introduction of the attention mechanism in this case significantly improves the accuracy of the grading model. By focusing on the most severe part of the burn area, the system is able to dynamically adjust the weights in the global grading to generate accurate grading results.

[0155] Embodiment 3:

[0156] A 55-year-old male patient suffered from a serious injury to the skin of his feet due to an accidental hot liquid burn. The doctor made a preliminary diagnosis and found that the patient had third-degree burns involving the epidermis, dermis, and part of the subcutaneous tissue. In order to further accurately assess the depth, area, and healing progress of the burn, the doctor decided to use the burn wound area image segmentation method of the present invention to assist in diagnosis and treatment planning. The patient's condition is shown in the figure. Figure 4 A is the healing process after skin autograft, and Figure 4 B shows a severe burn wound area without transplantation. This embodiment will explain in detail how to segment, grade and track the healing of a patient's foot burn wound by using hierarchical image segmentation technology combined with multispectral imaging, deep learning models and three-dimensional convolutional networks.

[0157] First, the doctor imaged the patient's foot burn area using a multispectral imaging device and an optical coherence tomography (OCT) device. The multispectral imaging device captured images of the burn area under visible, infrared, and ultraviolet light bands, and the OCT device provided three-dimensional depth data of the wound area. After the image acquisition was completed, the system performed preprocessing operations on the image, including image normalization, boundary detection, noise removal, and image enhancement processing to ensure the quality of the image data before segmentation.

[0158] Next, the system performs preliminary segmentation of the surface image of the burn area. After data preprocessing of the foot burn wound shown in Figure B, the system uses a convolutional neural network (CNN) to extract the features of the burn surface, such as color, texture, and edge changes. The system performs image convolution operations and extracts surface features using the following formula:

[0159] F θ (x,y)=∫ Ω k θ (x′,y′)·I(xx′,yy′)dx′dy′

[0160] Among them, F θ (x,y) represents the feature response at the image coordinate (x,y). Convolution kernel k θ(x′, y′) is adaptively adjusted to capture the surface changes of the burn area. Through this process, the system can automatically identify the edges of the burn and the severely damaged areas. For example, for Figure 4 In the area shown in B, the system detected that the color of the burn area was obviously reddish, the boundary was blurred, and the texture of the surface tissue showed rough and damaged characteristics, reflecting severe epidermal damage.

[0161] To further understand the depth characteristics of burns, especially Figure 4 The healing condition of the transplanted skin in A. The system uses a three-dimensional convolutional neural network (3DCNN) to process the collected three-dimensional depth data. Through the depth information obtained by the OCT device, the system can analyze the damage from the epidermis to the subcutaneous tissue.

[0162] The mathematical representation of the 3D convolution operation is as follows:

[0163] D θ (x,y,z)=∫ Γ k θ (x′,y′,z′)·I(xx′,yy′,zz′)dx′dy′dz′

[0164] D in the formula θ (x, y, z) reflects the three-dimensional feature response at the coordinate (x, y, z), and the z dimension represents the depth of the burn tissue. Through this operation, the system can accurately extract the deep-level injury characteristics of the burn area. Assume that the characteristic response value of the epidermis at a depth of z = 0.1 mm is 230, and at a depth of z = 1 mm in the dermis, the characteristic response value drops to 180, indicating that the dermis is more severely damaged, but the response value of the subcutaneous layer at z = 2 mm is only 120, reflecting relatively mild subcutaneous damage.

[0165] The system outputs these segmentation results in the form of images to provide doctors with a more detailed assessment of the depth of the injury. For example, the three-dimensional segmentation images generated by the system can clearly show the changes in the burn area at different depths, making it easier for doctors to judge the healing progress at different levels. In order to ensure the accuracy of the segmentation results, the present invention also introduces a feedback and correction mechanism. This mechanism uses the surface segmentation results as a priori input to dynamically adjust the segmentation boundaries of the deep tissue. The system corrects the deep segmentation results using the following formula:

[0166]

[0167] Among them, H n (R n-1 (x,y),λ n ) represents the correction operation for deep segmentation, R n-1 (x,y) is the result of surface segmentation, λ nRepresents the correction strength. Through this correction mechanism, the system can ensure that the segmentation boundaries of the surface and deep layers are consistent, avoiding the problem of cross-level boundary incoherence caused by segmentation errors.

[0168] After completing the multi-level segmentation, the system minimizes the energy function J(R n ,G) to optimize the segmentation results at different levels to ensure that the segmentation results at all levels can be merged into a complete burn wound segmentation map. This process is achieved through the following formula:

[0169]

[0170] Through this optimization process, the system can eliminate errors in the segmentation results at different levels and ensure that the segmentation boundaries are smooth and coherent. Subsequently, the system performs burn classification on the segmentation results and generates a global classification result by weighted fusion of surface features and depth features. The formula is as follows:

[0171]

[0172] Among them, w i is the feature weight coefficient, which adjusts the surface features and deep features For example, for severely burned areas of the epidermis, the system sets w1 = 0.7, while for less damaged areas of the dermis, w2 = 0.5. The resulting grading results provide doctors with accurate burn grade information, helping them develop personalized treatment plans.

[0173] The segmentation method of the present invention also has a real-time monitoring function. Figure 4 In the case of the transplanted skin in A, the system can track the skin healing process by regularly collecting images. By analyzing the segmentation results of consecutive images, the system can evaluate the healing progress and adjust the model parameters according to real-time changes. For example, when it detects that the epidermal area of ​​the healing area is gradually shrinking, the system will reduce the weight of the surface features to pay more attention to the healing of the deep tissue.

[0174] Through the burn wound area image segmentation method of the present invention, the system combines surface features and three-dimensional depth information to achieve accurate segmentation and grading of the foot burn area, and can track the healing progress in real time. Doctors can accurately evaluate the depth, area and healing status of the burn based on the segmented images and grading results generated by the system, and provide patients with more personalized and scientific treatment plans. This example demonstrates the potential for wide application of this method in burn diagnosis and treatment.

Claims

1. A method for image segmentation of a burn wound area, characterized in that The following steps are involved: S1. Adaptive spectral imaging segmentation based on optical properties of biological tissues: S1.

1. Collect burn wound images under different spectra using multispectral imaging equipment, including multiple bands of visible light, infrared light, and ultraviolet light; for each spectral image, use the optical properties of biological tissue to analyze the reflection and absorption characteristics of the burn area under different bands and mark potential burn areas; S1.

2. Design an adaptive algorithm to dynamically adjust the parameters of the imaging device including exposure and light intensity based on the feedback of different spectral features; enhance the characteristic contrast of the burn area; S1.

3. Preliminarily delineate the boundaries of the burn area based on the enhanced spectral image by threshold segmentation or region growing algorithm; S2. Burn tissue image segmentation based on hierarchical decomposition: S2.

1. Input the multispectral image into the hierarchical processing model to decompose different tissue layers of the burn area including the epidermis and dermis based on different bands or color information; S2.

2. On each layered result, use a feature extraction algorithm including texture analysis and edge detection to extract detailed features of each layer to distinguish burn tissues at different depths; S2.

3. Based on the hierarchical features, a convolutional neural network (CNN) architecture is used to independently segment each layer of burn tissue. The segmentation results of each layer are merged to generate a complete segmentation map of the burn wound area. S3. Boundary segmentation of the junction between deep burn wound and normal skin: S3.

1. In the segmentation results, identify the boundary area between the burn wound and the normal skin, and extract the boundary features of the area, including texture differences and optical reflection changes; S3.2, use GraphCut or LevelSet algorithm to optimize the local boundary of the junction; S4. Grading and segmentation of fusion burn depth and tissue structure: S4.

1. Combine optical coherence tomography (OCT) or ultrasound imaging to obtain three-dimensional depth information of the burn wound and form a preliminary depth distribution map; S4.

2. Design a deep learning model to combine the burn area segmentation results with the three-dimensional depth information to establish a burn grading model; classify the burn grades according to the depth characteristics, including epidermal burns and dermal burns; S5. Segmentation and tracking of healing process based on temporal features: S5.

1. During the burn healing process, regularly collect wound surface images and construct time series image data of the healing process; S5.

2. Analyze the image segmentation results at consecutive time points to extract the temporal characteristics of wound area reduction, boundary changes, and healing degree; S5.

3. Dynamically optimize the segmentation model based on the time series characteristics; adjust the segmentation parameters according to the healing progress.

2. The image segmentation method of the burn wound area according to claim 1, characterized in that The method for generating a complete burn wound area segmentation map comprises: The burn wound is regarded as a single multi-layer structure, each layer represents a different tissue structure including the epidermis, dermis, and subcutaneous layer; each layer presents a different feature pattern in the image; the unique features of each layer are extracted to design a multi-layer convolutional neural network architecture CNN, which processes the image layer by layer and designs different convolution kernels Extract organizational features at each level; the parameters of the convolution kernel θ n The process of tuning at different levels is expressed by the following formula: in, represents the feature response extracted at the nth level, (x, y) is the coordinate of the image on the two-dimensional plane; Ω represents the pixel space of the image, covering the entire image area in the integral; The convolution kernel function of the nth level, the convolution kernel parameter θ n It is an adjustable parameter corresponding to the structural features of each layer, which is used to extract the image features of this layer; I(xx′,yy′) is the pixel intensity value of the input image at the offset coordinate (xx′,yy′).

3. The image segmentation method of the burn wound area according to claim 2 is characterized in that The method for generating a complete burn wound area segmentation map includes: designing a feedback and correction mechanism, sharing part of the feature information, taking the surface segmentation result as the prior information of the next layer, and dynamically adjusting the segmentation strategy of the next layer; the feedback mechanism allows the segmentation processes of different levels to influence each other in boundary processing; setting the segmentation result of the previous layer, and correcting the deep segmentation result by extracting the gradient change boundary information.

4. The image segmentation method of the burn wound area according to claim 3 is characterized in that The method for generating a complete burn wound area segmentation map comprises: Define a global optimization objective function, and optimize based on the difference between the segmentation results of each level and the merged image; the goal of the optimization process is to eliminate boundary inconsistencies between levels by minimizing the energy function: Among them, J(R n ,G) is the global optimization energy function, and the multi-layer segmentation result is optimized by minimizing this function; R n (x, y) is the segmentation result of the nth level; G(x, y) is the global target segmentation result, which represents the desired complete segmented image after optimization; Ω represents the integration range, that is, the pixel space of the entire image.

5. The image segmentation method of the burn wound area according to claim 4 is characterized in that The method for generating a complete burn wound area segmentation map includes: combining three-dimensional depth data obtained by an optical coherence tomography (OCT) device; layered processing of burn tissue in three-dimensional space involves features of x and y dimensions, as well as changes in the depth dimension z; and extending the convolution operation to three-dimensional space to process tissue features of the burn wound at different depth layers.

6. The image segmentation method of the burn wound area according to claim 1, characterized in that The method for establishing a burn grading model comprises: Surface feature extraction is performed on the two-dimensional image of the burn area, including extracting local edge, color and texture information from the image to identify the burn boundary and changes in the wound morphology, as well as surface damage. The convolutional neural network calculates the response of the image through convolution operations, which is specifically expressed as: F θ (x,y)=∫ Ω k θ (x′,y′)·I(xx′,yy′)dx′dy′ in, F θ (x,y) represents the feature response at the coordinate (x,y), which is the feature extraction result of the convolution kernel on the input image; I(x,y) is the input two-dimensional burn image, where x and y are the horizontal and vertical coordinates of the image respectively; k θ (x′, y′) is the convolution kernel function, θ is the parameter of the convolution kernel, which is used to adaptively adjust the weight of the filter according to different levels to ensure that features of different scales can be extracted; Ω is the area of ​​image pixels, and the integral operation covers this area to extract features.

7. The image segmentation method of the burn wound area according to claim 6, characterized in that The method for establishing a burn grading model includes: combining three-dimensional depth information, analyzing the longitudinal depth changes of the burn area through a three-dimensional convolution network; extracting longitudinal features through three-dimensional convolution to classify burns at different depth levels; and using the three-dimensional convolution operation to extract depth information of the burn area.

8. The image segmentation method of the burn wound area according to claim 7, characterized in that The method for establishing a burn grading model comprises: The attention mechanism is introduced to focus on the serious part of the burn area, so as to extract and grade the features of the local area. Through the attention mechanism, the focus of the model is dynamically adjusted according to the depth characteristics of the burn area. In the global grading process, the surface and depth features are weighted and fused through the following formula to achieve grading judgment: Among them, G(F,D) is the global classification result, which means that based on surface features and deep features The comprehensive weighted result of It is the surface feature response of the two-dimensional image at the i-th level, reflecting the information of the burn surface; is the depth feature response of the three-dimensional image at the i-th level, reflecting the longitudinal information of the burn depth; w i is the feature weight coefficient, which is used to adjust the importance of surface features and depth features; w i The value of is adjusted through model learning; N is the total number of model levels, indicating the number of fusions at different feature levels.

Images