Skin wound surface area measurement grading method and system based on deep learning

Abstract

The invention provides a skin wound area measuring and grading method and system based on deep learning, and relates to the technical field of skin wound area measuring and grading, and the method comprises the steps: collecting initial visible light and near-infrared light band images of a skin wound of a target patient, respectively enhancing the images, and fusing the images to obtain a skin wound image; segmenting the image by using a semantic segmentation algorithm to obtain a binary image, and counting the number of wound surface pixels to obtain an actual wound surface area; and then the texture roughness and color uniformity characteristics of the wound surface are extracted through deep learning, the wound surface is graded in combination with a clinical grading standard, and the wound surface grading result of the patient is obtained, so that accurate measurement of the skin wound surface area and grading evaluation based on the texture and color characteristics can be realized, and a reliable basis is provided for clinical wound surface monitoring.

Description

Technical Field

[0001] This application relates to the technical field of skin wound area measurement and grading, and in particular to a deep learning-based method and system for skin wound area measurement and grading. Background Technology

[0002] In outpatient treatment, inpatient nursing, and postoperative recovery monitoring scenarios at all levels of medical institutions, the size and healing status of skin wounds are the core basis for assessing the condition, formulating treatment plans, and adjusting nursing strategies. At the same time, home self-monitoring by patients with chronic wounds also requires convenient and accurate measurement and grading methods to track wound recovery progress in real time and provide timely feedback to medical staff. Therefore, there is an urgent need for a skin wound area measurement and grading technology that can balance accuracy and practicality to meet the dual needs of efficient clinical diagnosis and treatment and patient home monitoring.

[0003] Currently, the measurement and grading of skin wounds are mainly achieved through two methods: one is the manual measurement method, in which medical staff use tools such as rulers and grid paper to directly measure the length and width of the wound, estimate the area based on experience, and then subjectively judge the grading result based on the appearance of the wound and the amount of exudation; the other is the basic machine-assisted method, which acquires a single visible light image of the wound, uses traditional image processing techniques such as threshold segmentation to distinguish the wound from the background area, calculates the number of pixels to convert the area, and the grading is mainly completed based on the area threshold or a single texture feature.

[0004] However, existing technologies have drawbacks, the most significant being insufficient measurement accuracy and comprehensiveness of grading: manual measurement is easily affected by operating techniques and subjective experience, making it difficult to accurately capture the true extent of irregular wounds, resulting in poor consistency of grading results; basic machine-assisted methods rely only on single-band images, failing to fully acquire information about deep tissues in the wound, making boundary segmentation susceptible to environmental interference, and grading relies only on a single indicator, failing to comprehensively reflect the wound healing status, and thus failing to meet the clinical needs for precise and comprehensive diagnostic and treatment assessment. Summary of the Invention

[0005] The purpose of this application is to provide a deep learning-based method and system for measuring and grading skin wound area, in order to solve the problem that existing technologies cannot meet the clinical needs for precise and comprehensive diagnostic and treatment assessment.

[0006] To address the aforementioned technical problems, in a first aspect, this application provides a deep learning-based method for measuring and grading skin wound area, comprising: Acquire initial visible light and initial near-infrared light images of the target patient's skin wound; Image enhancement processing is performed on the initial visible light image and the initial near-infrared light image respectively to obtain the target visible light image and the target near-infrared light image. The target visible light image and the target near-infrared light image are then fused to obtain the skin wound image. A semantic segmentation algorithm is used to segment the skin wound image to obtain a binarized image. The number of pixels belonging to the skin wound in the binarized image is calculated to obtain the actual wound area. The skin wound image is subjected to feature extraction processing based on deep learning to obtain the texture roughness and color uniformity features of the skin wound. Based on the actual wound area, the texture roughness, and the color uniformity, and combined with the preset clinical skin wound grading standards, the skin wound is graded to obtain the skin wound grading results for the target patient.

[0007] Optionally, the step of using a semantic segmentation algorithm to segment the skin wound image to obtain a binarized image, and calculating the number of pixels belonging to the skin wound in the binarized image to obtain the actual wound area, includes: The skin wound image is processed to obtain multi-scale depth features. Based on the multi-scale depth features, a semantic segmentation algorithm is used to obtain the boundary contour between the wound area and the normal skin area in the skin wound image. Based on the boundary contour, the pixels in the skin wound image are divided into wound region pixels and background region pixels, and a preliminary segmentation image is generated based on the wound region pixels and the background region pixels; The preliminary segmented image is subjected to contour optimization processing to obtain a binarized image, and the total number of pixels in the wound area in the binarized image is counted. Based on the total number of pixels in the wound area and the preset actual size of the unit pixel, the actual wound area of ​​the skin wound is obtained.

[0008] Optionally, the step of extracting and processing the skin wound image to obtain multi-scale depth features, and using a semantic segmentation algorithm based on the multi-scale depth features to obtain the boundary contour between the wound area and the normal skin area in the skin wound image, includes: Multiple processing layers with different receptive fields are used to extract layered features from the skin wound image, resulting in shallow features containing the fine structure of the wound and deep features containing the overall morphology of the wound. The shallow features and the deep features are fused to obtain multi-scale depth features, which are then used to construct a feature map. Based on the multi-scale depth features, a semantic segmentation algorithm with an edge-aware feature enhancement module is used to perform pixel-level classification of the skin wound image. During the pixel-level classification process, the edge-aware feature enhancement module calculates the gradient information of the feature map to enhance the feature response at the junction of wound tissue and normal skin, thereby obtaining the pixels of wound tissue and normal skin tissue. Based on the pixels of the wound tissue and the pixels of the normal skin tissue, the boundary point between the wound area and the normal skin area is obtained. Adjacent boundary points are connected to form the boundary contour between the wound area and the normal skin area.

[0009] Optionally, the step of performing contour optimization processing on the initially segmented image to obtain a binarized image, and counting the total number of pixels in the wound region of the binarized image, includes: Connectivity analysis is performed on the preliminary segmented image to obtain discontinuous regions between boundary points; Based on the near-infrared band information and visible light color information in the skin wound image, the tissue characteristics of the discontinuous region are determined. According to the tissue characteristics, an adaptive interpolation algorithm is used to complete the boundary in order to establish a complete closed boundary contour. The closed boundary contour is smoothed and filtered to obtain a binarized image; In the binarized image, the connected pixel groups of the wound area are identified, and the total number of pixels contained in the connected pixel groups is counted to obtain the total number of pixels in the wound area.

[0010] Optionally, the step of performing deep learning-based feature extraction processing on the skin wound image to obtain the texture roughness and color uniformity features of the skin wound includes: Based on the binarized image, the wound area image is extracted from the skin wound image; The wound area image is input into a pre-trained feature extraction network, which includes parallel texture feature branches and color feature branches; In the texture feature branch, multiple Gabor filter layers and the first convolutional layer are used to extract deep texture features of the wound area. Based on the deep texture features of the wound area, the gray value variation amplitude between adjacent pixels in the wound area image is calculated to obtain the texture roughness. In the color feature branch, the color distribution is learned through the second convolutional layer, and based on the color distribution, the concentration of different color channel values ​​in the wound area image is statistically analyzed to obtain the color uniformity feature.

[0011] Optionally, the step of grading the skin wound based on the actual wound area, the texture roughness, and the color uniformity, combined with a preset clinical skin wound grading standard, to obtain the skin wound grading result for the target patient includes: The actual wound area is matched with the area threshold in the preset clinical skin wound grading standard to obtain the area grading parameter. The texture roughness is compared with the texture reference feature in the preset clinical skin wound grading standard to obtain the texture grading parameter. The color uniformity feature is analyzed with the color distribution benchmark in the preset clinical skin wound grading standard to obtain the color grading parameter. The area grading parameters, texture grading parameters, and color grading parameters are input into a preset multi-dimensional grading decision model. The parameters are then weighted and fused through the preset multi-dimensional grading decision model to obtain a comprehensive evaluation value. The comprehensive assessment value is compared with the grading threshold in the preset clinical skin wound grading standard to obtain the skin wound grading result of the target patient.

[0012] Optionally, the step of performing image enhancement processing on the initial visible light image and the initial near-infrared light image respectively to obtain a target visible light image and a target near-infrared light image, and fusing the target visible light image and the target near-infrared light image to obtain a skin wound image, includes: The initial visible light image and the initial near-infrared light image are subjected to orientation transformation and brightness adjustment operations, respectively, to obtain the target visible light image and the target near-infrared light image; Based on the target visible light image, the edge contour information of the skin wound is obtained; based on the target near-infrared light image, the tissue depth information of the skin wound is obtained. The edge contour information and the tissue depth information are combined to obtain a skin wound image.

[0013] Secondly, this application provides a deep learning-based skin wound area measurement and grading system, comprising: The acquisition module is used to acquire initial visible light and initial near-infrared light images of the target patient's skin wound. The fusion module is used to perform image enhancement processing on the initial visible light image and the initial near-infrared light image respectively to obtain the target visible light image and the target near-infrared light image, and to fuse the target visible light image and the target near-infrared light image to obtain the skin wound image; The segmentation module is used to segment the skin wound image using a semantic segmentation algorithm to obtain a binarized image, and to calculate the number of pixels belonging to the skin wound in the binarized image to obtain the actual wound area. The extraction module is used to perform deep learning-based feature extraction processing on the skin wound image to obtain the texture roughness and color uniformity features of the skin wound. The grading module is used to grade the skin wound based on the actual wound area, the texture roughness, and the color uniformity, combined with a preset clinical skin wound grading standard, to obtain the skin wound grading result of the target patient.

[0014] Thirdly, this application provides an electronic device, comprising: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the deep learning-based skin wound area measurement and grading method as described in the first aspect above.

[0015] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the deep learning-based skin wound area measurement and grading method described in the first aspect above.

[0016] The technical solution of this application has the following beneficial effects: This application first acquires visual appearance information and deep tissue correlation information of the wound simultaneously to provide comprehensive data support for subsequent multi-dimensional analysis, avoiding the limitations of incomplete information from single-band images. Next, enhancement processing improves image clarity and information recognition, and fusion operations integrate dual-band core features, making information such as wound edges and tissue state more prominent, thus laying a high-quality data foundation for subsequent segmentation and feature extraction. Then, semantic segmentation algorithms are used to accurately distinguish the wound from the background and obtain a binarized image, clearly defining the wound area. Pixel statistics and size conversion are then used to objectively quantify the wound area, improving the accuracy and consistency of area measurement. Subsequently, deep learning technology is used to accurately capture key surface features of the wound, obtaining core indicators reflecting the healing status, thus providing multi-dimensional assessment criteria for grading in addition to area, further enriching the assessment dimensions. Finally, integrating three core features—area, texture, and color—and comparing them with clinical standards for comprehensive grading, the grading results more closely reflect the actual healing status of the wound, providing a comprehensive and reliable reference for treatment planning.

[0017] Furthermore, this application improves the accuracy of wound boundary contour recognition through multi-scale depth features, enhances the integrity and regularity of the wound area through contour optimization processing, and achieves accurate quantification of wound area by combining pixel statistics and size conversion, thereby further improving the reliability and accuracy of area measurement and providing more accurate quantitative data support for wound assessment in clinical diagnosis and treatment. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating a deep learning-based method for measuring and grading skin wound area, provided as an embodiment of this application; Figure 2 A schematic diagram illustrating a specific implementation of a deep learning-based method for measuring and grading skin wound area, as provided in this application embodiment; Figure 3 A schematic diagram of a deep learning-based skin wound area measurement and grading system provided in this application embodiment; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0020] In current skin wound measurement and grading methods, manual measurements are easily influenced by subjective experience, resulting in insufficient accuracy in estimating the area of ​​irregular wounds and poor consistency in grading results. Basic machine-assisted methods rely on single-band images, making it difficult to obtain information about deep tissues within the wound. This leads to boundary segmentation being easily affected by environmental interference, and the grading is based on only a single indicator, failing to comprehensively reflect the wound healing status. These problems make it difficult for existing technologies to meet the needs of clinical diagnosis and treatment for precise and comprehensive assessment, and also cause inconvenience for patients monitoring at home.

[0021] To address the aforementioned issues, this application provides a deep learning-based method for measuring and grading skin wound area. This method simultaneously acquires visible and near-infrared images of the wound, enhances and fuses them to obtain a fused image containing comprehensive information, and then uses a semantic segmentation algorithm to accurately distinguish the wound from the background area to calculate the actual area. Simultaneously, it extracts the texture and color features of the wound and combines them with clinical standards for comprehensive grading. This approach not only compensates for the insufficient information and low measurement accuracy of a single image through dual-band fusion and precise segmentation, but also solves the problem of one-sided assessment through multi-dimensional feature grading, achieving precise and comprehensive wound measurement and grading, providing reliable support for clinical diagnosis and treatment and home monitoring of patients.

[0022] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] The core of this application is to provide a deep learning-based method for measuring and grading skin wound area, and a flowchart of one specific implementation is shown below. Figure 1 As shown, the method includes: S101. Acquire initial visible light and initial near-infrared light images of the target patient's skin wound.

[0024] In the above scheme, the skin wound refers to the area of ​​skin damage on the target patient's body surface caused by trauma, surgery or disease. The initial visible light image refers to the original image of the skin wound directly acquired without any processing in the 400-760nm range of wavelengths perceptible to the human eye. This image can present the appearance color, intuitive surface morphology and visual boundary with the surrounding normal skin of the wound. The initial near-infrared light image refers to the original image of the skin wound directly acquired without any processing in the 760-2500nm range of wavelengths invisible to the human eye. This image can penetrate the skin surface and capture relevant information of the tissue beneath the wound.

[0025] In this application example, the image acquisition and processing involves: first, preparing a suitable multispectral acquisition device and adjusting the wound posture of the target patient to ensure that the acquisition device lens is aimed at the skin wound without obstruction, so that the wound is completely centered in the acquisition field of view, thereby avoiding interference from surrounding debris; then, activating the visible light band acquisition function of the acquisition device and taking pictures of the skin wound according to the device's preset acquisition parameters such as resolution and exposure time, which can obtain an initial visible light image that presents the appearance, color, surface morphology, and visual boundaries of the wound; finally, keeping the relative position of the acquisition device and the wound unchanged, switching to the near-infrared band acquisition function, and taking pictures again using the same preset acquisition parameters, which can obtain an initial near-infrared light image that penetrates the skin surface and captures information related to deep tissues.

[0026] In this embodiment of the invention, by simultaneously acquiring two different bands of original images, the visual information of the appearance of the skin wound and the information related to the deep tissues are completely obtained. This can make up for the limitation that a single band image can only capture partial information, and can provide comprehensive and undisturbed basic data for subsequent image enhancement and fusion processing.

[0027] S102. Perform image enhancement processing on the initial visible light image and the initial near-infrared light image respectively to obtain the target visible light image and the target near-infrared light image. Then fuse the target visible light image and the target near-infrared light image to obtain the skin wound image.

[0028] In one specific implementation, step S102 includes: Step 1021: Perform orientation transformation and brightness adjustment operations on the initial visible light image and the initial near-infrared light image respectively to obtain the target visible light image and the target near-infrared light image.

[0029] Orientation transformation refers to adjusting the angle and position of the image to make it more regular, which can be used to eliminate image tilt or offset caused by posture deviation during the acquisition process; brightness adjustment refers to optimizing the brightness and contrast of the image, which can be used to improve problems such as overly dark, overexposed, or unclear details in the image.

[0030] In this application example, image enhancement processing is performed. First, in step 1021, an image rotation and translation algorithm is used to perform orientation transformation on the initial visible light image and the initial near-infrared light image respectively. The image tilt is corrected according to the edge reference line of normal skin in the image to eliminate the positional deviation during acquisition. Then, histogram equalization technology is used to adjust the brightness and contrast of the two images respectively to optimize the brightness distribution and contrast of the images, which can highlight the details of the wound area. Finally, the target visible light image and the target near-infrared light image are obtained.

[0031] For example, when medical staff at a medical institution target a skin wound on a patient's back, they first perform an orientation transformation on the initial visible light image and then use the system to identify the horizontal edge line of normal skin in the image to obtain an image tilt angle of 3 degrees. Subsequently, based on this angle, they perform reverse rotation correction on the image to eliminate the tilt deviation. They then statistically analyze the image grayscale value distribution range of 40 to 200 and redistribute the grayscale values ​​using a histogram equalization algorithm to make the image brightness uniform, thus completing the brightness adjustment.

[0032] Step 1022: Based on the target visible light image, obtain the edge contour information of the skin wound; based on the target near-infrared light image, obtain the tissue depth information of the skin wound.

[0033] Among them, edge contour information refers to the characteristic data that characterizes the shape and position of the boundary between the wound and normal skin. For example, edge contour information includes boundary point coordinates, contour direction, etc.; tissue depth information refers to the relevant data that reflects the tissue layers and thickness below the wound. For example, tissue depth information can be obtained based on the near-infrared light penetration characteristics.

[0034] In step 1022, edge contour information of the skin wound is extracted based on the target visible light image, and Gaussian filtering is used to reduce noise caused by ambient light interference. Then, the Canny edge detection algorithm is used to calculate the gray-level gradient of pixels in the image to determine potential boundaries. After that, a double threshold is set to filter effective boundary points and remove false boundaries and discontinuous boundary segments, finally forming complete wound edge contour data. At the same time, tissue depth information of the skin wound is extracted based on the target near-infrared light image, and the gray value mapping technique is used to establish the correspondence between image gray values ​​and tissue depth by taking advantage of the characteristic that the intensity of reflected light from tissues at different depths after near-infrared light penetrates the skin. Then, by traversing the gray values ​​of each wound pixel in the image and combining the correspondence transformation, the tissue depth data corresponding to each pixel is obtained.

[0035] For example, the edge contour information of the target visible light image is extracted by the Canny edge detection algorithm built into the system. First, the noise is reduced by a Gaussian filter with a standard deviation of 1.4. Then, the gray-level gradients in the x and y directions of the image are calculated. Subsequently, a high threshold of 80 and a low threshold of 40 are set, and effective boundary points are obtained by filtering according to the thresholds. Then, all effective boundary points are connected in sequence to form complete wound edge contour coordinate data.

[0036] Subsequently, the tissue depth information of the target near-infrared image was extracted using grayscale mapping technology. First, all pixels in the wound area were traversed, and one pixel with a grayscale value of 120 was selected. Then, a mapping formula was established. Where D represents tissue depth, G represents pixel grayscale value, k represents scaling factor 0.02, and b represents correction constant 0, the corresponding tissue depth is calculated to be 2.4 by substituting into the formula, and the tissue depth conversion of all wound pixels is completed in the same way to form complete tissue depth data.

[0037] Step 1023: Combine the edge contour information and the tissue depth information to obtain a skin wound image.

[0038] Among them, skin wound images refer to comprehensive images that integrate edge contour information and tissue depth information, containing both the appearance of the wound and the characteristics of deep tissues, which can be used for subsequent precise analysis.

[0039] In step 1023, feature stitching technology can be used to integrate the extracted edge contour information and tissue depth information, and the coordinate data of the edge contour can be matched with the tissue depth data at the corresponding location to generate a skin wound image containing the wound appearance boundary and deep tissue features.

[0040] For example, the feature stitching function of an image processing system can be used to associate and integrate edge contour coordinate data with tissue depth data at the corresponding location to generate a skin wound image.

[0041] This application optimizes the quality of the original image through orientation transformation and brightness adjustment, eliminates deviations in the acquisition process and highlights key information, and then obtains two core information types, the edge contour and tissue depth of the wound, through precise feature extraction technology. Finally, a comprehensive skin wound image is generated by feature combination.

[0042] S103. Using a semantic segmentation algorithm, the skin wound image is segmented to obtain a binarized image. The number of pixels belonging to the skin wound in the binarized image is calculated to obtain the actual wound area.

[0043] In one specific implementation, such as Figure 2 As shown, step S103 includes: Step 1031: Extract and process the skin wound image to obtain multi-scale depth features. Based on the multi-scale depth features, use a semantic segmentation algorithm to obtain the boundary contour between the wound area and the normal skin area in the skin wound image.

[0044] Among them, semantic segmentation algorithm refers to deep learning technology that performs pixel-level category division of images, which can be used to accurately distinguish different target regions in an image; multi-scale deep features are obtained by performing layered feature extraction on skin wound images through multiple processing layers with different receptive fields. Among them, the processing layer with a smaller receptive field can extract shallow features that reflect the fine structure of the wound, while the processing layer with a larger receptive field can extract deep features that reflect the overall morphology of the wound. Furthermore, the receptive field refers to the range of perception of the input image by the image processing layer, where different receptive field processing layers can extract different scale features; shallow features refer to feature data containing fine structures extracted from the image surface, deep features refer to abstract feature data containing the overall shape obtained through multi-layer processing, and feature maps refer to the visualization or digital carriers that carry feature data.

[0045] Step 1031 may specifically include the following steps: using multiple processing layers with different receptive fields to perform layered feature extraction on the skin wound image, obtaining shallow features containing the fine structure of the wound and deep features containing the overall morphology of the wound; fusing the shallow features and the deep features to obtain multi-scale depth features to construct a feature map; based on the multi-scale depth features, using a semantic segmentation algorithm with an edge-aware feature enhancement module to perform pixel-level classification on the skin wound image, in which the edge-aware feature enhancement module calculates the gradient information of the feature map to enhance the feature response at the junction of wound tissue and normal skin, obtaining wound tissue pixels and normal skin tissue pixels; based on the wound tissue pixels and normal skin tissue pixels, obtaining the boundary point between the wound area and the normal skin area, connecting adjacent boundary points to form the boundary contour between the wound area and the normal skin area.

[0046] Among them, the edge-aware feature enhancement module refers to the auxiliary module in the semantic segmentation algorithm, which can be used to enhance the feature response at the boundary between the target and the background; the wound tissue pixel refers to the image pixel determined as the wound area, the normal skin tissue pixel refers to the image pixel determined as the normal skin area, the dividing point refers to the pixel boundary position between the wound area and the normal skin area, and the boundary contour refers to the closed line representing the wound range formed by connecting all the dividing points.

[0047] For example, multi-scale depth features are extracted through multiple convolutional processing layers. The shallow convolutional layer captures the small depressions and edge wrinkles of the wound on the patient's back to obtain shallow features, while the deep convolutional layer captures the irregular overall shape of the wound to obtain deep features. The two features are then fused to form a multi-scale depth feature map. Subsequently, the gradient information of the feature map is calculated by a semantic segmentation algorithm with an edge-aware feature enhancement module to enhance the feature response at the junction of the wound and normal skin. The wound tissue pixels and normal skin tissue pixels are determined pixel by pixel, and the boundary points are connected to form the boundary contour.

[0048] Step 1032: Based on the boundary contour, the pixels in the skin wound image are divided into wound area pixels and background area pixels, and a preliminary segmentation image is generated based on the wound area pixels and the background area pixels.

[0049] In step 1032, all pixels in the skin wound image are classified based on the boundary contour. For example, pixels inside and on the boundary contour are identified as wound area pixels, and the area outside the contour is identified as background area pixels. The background area pixels include pixels in non-skin areas and pixels in normal skin areas.

[0050] Step 1033: Perform contour optimization processing on the preliminary segmented image to obtain a binarized image, and count the total number of pixels in the wound area of ​​the binarized image.

[0051] Among them, contour optimization processing refers to a series of operations to complete and smooth the boundaries of the initially segmented image; binarized image refers to an image that distinguishes the wound area from the non-wound area using only two pixel values; and the total number of pixels in the wound area refers to the number of pixels belonging to the wound area in the binarized image.

[0052] In one specific implementation, step 1033 may specifically include the following steps: performing connectivity analysis on the preliminary segmented image to obtain discontinuous regions between boundary points; determining the tissue characteristics of the discontinuous regions based on the near-infrared band information and visible light color information in the skin wound image; using an adaptive interpolation algorithm to complete the boundaries according to the tissue characteristics to establish a complete closed boundary contour; performing smoothing filtering on the closed boundary contour to obtain a binarized image; identifying connected pixel groups of the wound region in the binarized image, counting the total number of pixels contained in the connected pixel groups, and obtaining the total number of pixels in the wound region.

[0053] Among them, connectivity analysis refers to the technique of detecting the connectivity of pixels in an image, which can be used to identify discontinuous boundary regions; adaptive interpolation algorithm refers to the completion technique of dynamically adjusting the interpolation method according to the characteristics of the region; closed boundary contour refers to the complete and uninterrupted boundary lines of the wound region; and smoothing filtering refers to the operation of eliminating boundary jaggedness and optimizing boundary morphology.

[0054] For example, connectivity analysis is performed on the preliminary segmented image, and three discontinuous boundary regions are identified. Then, by combining the near-infrared tissue depth information and visible light color information in the fused image, it is determined that these regions are all extension regions of the wound. An adaptive interpolation algorithm is then used to complete the boundary points to form a closed boundary contour. After smoothing and filtering, a binarized image is obtained. Finally, the total number of connected pixel groups in the wound region of the image is counted to be 15,000.

[0055] Step 1034: Based on the total number of pixels in the wound area and combined with the preset actual size of the unit pixel, obtain the actual wound area of ​​the skin wound.

[0056] Among them, the actual size of a unit pixel refers to the real physical size corresponding to a single pixel in the preset image, which can be used to convert the number of pixels into the actual area; the actual wound area refers to the real physical area of ​​the wound calculated by the number of pixels and the actual size of a unit pixel.

[0057] In step 1034, the actual size of a unit pixel is first determined by geometric conversion based on the hardware parameters of the image acquisition device and the shooting distance during acquisition, that is, the physical area corresponding to a single pixel in the real scene. Then, the total number of pixels in the wound area obtained by statistics is multiplied by the actual size of the unit pixel to obtain the actual wound area of ​​the skin wound.

[0058] For example, if the pixel pitch of the image sensor of the acquisition device is set to 0.0002 cm, the shooting distance is 18 cm, and the lens focal length is 9 cm, then the actual size of the unit pixel can be converted using the formula... Calculate the actual size of a unit pixel, where S represents the actual size of a unit pixel, d represents the shooting distance, P represents the pixel pitch, and L represents the lens focal length. Substitute the values ​​into the formula to calculate S as 0.0004 square centimeters. Then multiply the total number of pixels in the wound area, 15000, by the actual size of a unit pixel, 0.0004 square centimeters to obtain the actual wound area as 6 square centimeters.

[0059] This application achieves accurate differentiation between wound areas and other areas by combining multi-scale depth features with enhanced semantic segmentation algorithms. Contour optimization further improves the regularity and accuracy of the boundaries, ensuring the reliability of pixel statistics in the wound area. Then, through scientific conversion logic, the number of pixels is converted into the actual wound area, enabling objective quantification of the wound area. This overcomes the shortcomings of traditional measurement methods, such as insufficient accuracy and strong subjectivity, and provides accurate and reliable quantitative data for subsequent grading based on multi-dimensional features, further ensuring the scientific nature and practicality of the final grading results.

[0060] S104. Perform deep learning-based feature extraction processing on the skin wound image to obtain the texture roughness and color uniformity features of the skin wound.

[0061] In this step, deep learning-based feature extraction processing refers to the technique of automatically identifying and extracting key features from images using a trained neural network model.

[0062] In one specific implementation, step S104 includes: Step 1041: Based on the binarized image, extract the wound area image from the skin wound image.

[0063] Among them, the wound area image refers to a partial image that is cropped from a skin wound image and contains only the wound area.

[0064] In step 1041, the pixel range of the wound area is determined based on the binarized image. Then, the characteristic that only the pixels of the wound area have specific values ​​in the binarized image is used to accurately extract the corresponding pixel range from the skin wound image to obtain a wound area image that only contains the wound area. This can completely eliminate the interference of normal skin and non-skin areas on feature extraction.

[0065] For example, based on the marked range of the pixels in the wound area of ​​the binarized image, the corresponding area is first extracted from the fused image, and a wound area image with a size of 500×500 pixels is obtained, and the total number of pixels in the image is counted as 500×500=250000.

[0066] Step 1042: Input the wound area image into a pre-trained feature extraction network, which includes parallel texture feature branches and color feature branches.

[0067] Among them, the pre-trained feature extraction network refers to a neural network model that has the ability to recognize features after being trained on a large amount of image data, and the parallel texture feature branch and color feature branch refer to two independent processing channels that run in parallel in the network, which are used to extract texture features and color features respectively. Step 1043: In the texture feature branch, multiple Gabor filter layers and the first convolutional layer are used to extract the deep texture features of the wound area. Based on the deep texture features of the wound area, the gray value change amplitude between adjacent pixels in the wound area image is calculated to obtain the texture roughness.

[0068] Among them, the Gabor filter layer refers to the image processing layer that is sensitive to texture in different directions and is used to capture multi-directional texture information; the first convolutional layer refers to the network layer in the texture feature branch that is used to deepen the extraction of texture information; the deep texture feature refers to the feature data that can reflect the essence of the wound texture after multi-layer processing; the gray value change amplitude refers to the degree of difference in gray value between adjacent pixels and is used to characterize the texture roughness; the texture roughness refers to the feature index that describes the unevenness, smoothness or roughness of the wound surface.

[0069] For example, the image of the wound area is input into a pre-trained feature extraction network. The texture feature branch of this network first activates eight Gabor filter layers in different directions for filtering, and then outputs deep texture features through the first convolutional layer. Based on these features, a formula is applied. Calculate the magnitude of grayscale value change between adjacent pixels, where, This indicates the magnitude of the change in grayscale value between adjacent pixels. This represents the grayscale value of the i-th pixel. This represents the grayscale value of the j-th pixel adjacent to the i-th pixel, and is calculated by selecting 100 consecutive pairs of adjacent pixels from the top left corner of the wound area image. For example... For 120, When it is 115, then ; For 115, When it is 130, then After performing calculations on 100 pairs of pixels in sequence, 100 results are obtained. The values ​​are summed to get 820, and then the mean is calculated, which is 820 ÷ 100 = 8.2. This mean is the texture roughness.

[0070] Step 1044: In the color feature branch, the color distribution is learned through the second convolutional layer, and based on the color distribution, the concentration of different color channel values ​​in the wound area image is statistically analyzed to obtain the color uniformity feature.

[0071] Among them, the second convolutional layer refers to the network layer in the color feature branch used to learn the color distribution pattern. Color distribution refers to the distribution of different color channel values ​​in the wound area image. Color channel value refers to the quantitative data representing color in the image. Concentration refers to the aggregation or dispersion state of color channel values. Color uniformity feature is a feature index describing whether the color of the wound area is uniform.

[0072] For example, the network's color feature branch learns the color distribution through the second convolutional layer and uses the variance formula. Calculate the concentration of color channel values, where, Let n represent the variance, and n represent the total number of pixels in the wound area image, for example, 250,000. This represents the color channel value of the k-th pixel. This represents the mean of all pixel color channel values. For example, summing all pixel color channel values ​​yields 30,000,000. First, calculate... 30,000,000 ÷ 250,000 = 120, then calculate each... The squared difference between 120 and 2500000 is summed to get 2500000. Finally, 2500000 is divided by 250000 to get 10, which means 10 is the color uniformity characteristic.

[0073] This application uses binarized images to accurately locate the wound area and eliminate interference from irrelevant areas. Then, it uses a pre-trained bi-branch feature extraction network to simultaneously and accurately extract the texture roughness and color uniformity features of the wound. This fully utilizes the feature recognition advantages of deep learning and avoids the subjectivity and limitations of manual feature extraction.

[0074] S105. Based on the actual wound area, the texture roughness, and the color uniformity, and in conjunction with the preset clinical skin wound grading standards, the skin wound is graded to obtain the skin wound grading result of the target patient.

[0075] In one specific implementation, step S105 includes: Step 1051: Match the actual wound area with the area threshold in the preset clinical skin wound grading standard to obtain area grading parameters; compare the texture roughness with the texture reference features in the preset clinical skin wound grading standard to obtain texture grading parameters; and perform dispersion analysis on the color uniformity feature with the color distribution benchmark in the preset clinical skin wound grading standard to obtain color grading parameters.

[0076] Among them, the pre-defined clinical skin wound grading standard refers to the assessment standard based on medical clinical experience and research, which includes area thresholds, texture reference features and color distribution benchmarks. Area threshold refers to the numerical limit for dividing area levels in the grading standard; texture reference feature refers to the texture feature template corresponding to different healing states in the grading standard; color distribution benchmark refers to the color distribution reference for normal and abnormal wounds in the grading standard. Area grading parameter refers to the grade quantification value obtained after matching the actual wound area with the area threshold; texture grading parameter refers to the grade quantification value after comparing the texture roughness with the reference feature; color grading parameter refers to the grade quantification value obtained after dispersion analysis.

[0077] In step 1051, the actual wound area, texture roughness, and color uniformity features are quantified and graded. The actual wound area is then matched with the area threshold in the preset clinical skin wound grading standard, and the corresponding area grading parameters are output according to the threshold interval to which the actual area belongs. At the same time, the texture roughness is compared with the texture reference features in the grading standard, and the corresponding texture grading parameters are determined by calculating the feature similarity between the two. Then, the color uniformity features are analyzed for dispersion with the color distribution benchmark in the grading standard, and the corresponding color grading parameters are output according to the degree of difference between the features and the benchmark.

[0078] For example, in the clinical skin wound grading standard, the area threshold is divided into parameter 1 for 0 to 5 square centimeters, parameter 2 for 5 to 10 square centimeters, and parameter 3 for more than 10 square centimeters. Since the actual area of ​​the patient's back wound is 6 square centimeters, which falls within the range of 5 to 10 square centimeters, the area grading parameter 2 is obtained. Then, the texture reference feature is divided into parameters 1 to 5 for roughness, parameter 2 for 5 to 10, and parameter 3 for more than 10. Since the texture roughness is 8.2, which falls within the range of 5 to 10, the texture grading parameter 2 is obtained. Subsequently, the variance threshold of the color distribution benchmark is divided into parameters 1 for less than 5, 2 for 5 to 15, and 3 for more than 15. By using dispersion analysis to determine that the color uniformity feature 10 is in the range of 5 to 15, the color grading parameter 2 is obtained.

[0079] Step 1052: Input the area grading parameters, texture grading parameters and color grading parameters into a preset multi-dimensional grading decision model, and perform parameter weighted fusion through the preset multi-dimensional grading decision model to obtain a comprehensive evaluation value.

[0080] Among them, the multi-dimensional hierarchical decision-making model refers to an intelligent evaluation model that can integrate multiple types of parameters after training, and the hierarchical threshold refers to the comprehensive value boundary for classifying wound levels in the hierarchical standard.

[0081] In step 1052, the area grading parameters, texture grading parameters, and color grading parameters are input into a preset multi-dimensional grading decision model. The model first assigns different weights to the three grading parameters of area, texture, and color based on clinical diagnosis and treatment experience. Since the area parameter is directly related to the severity of the wound, the area parameter has a relatively higher weight. Then, the three parameters are calculated using a weighted fusion algorithm to obtain a comprehensive evaluation value that can comprehensively reflect the wound status.

[0082] For example, in a multi-dimensional hierarchical decision-making model, the weights of the three types of parameters are set as area 0.4, texture 0.3, and color 0.3 respectively, and a weighted fusion formula is used. Where S represents the overall evaluation value, Indicates the area parameter weight. Indicates the texture parameter weights. Indicates the weight of the color parameter. Indicates area grading parameters, Indicates texture grading parameters, The value represents the color grading parameter. Substituting the values, we get S = 0.4 × 2 + 0.3 × 2 + 0.3 × 2 = 2.0.

[0083] Step 1053: Compare the comprehensive evaluation value with the grading threshold in the preset clinical skin wound grading standard to obtain the skin wound grading result of the target patient.

[0084] In step 1053, the comprehensive assessment value is compared with the grading threshold in the preset clinical grading standard, and the threshold range to which the comprehensive assessment value belongs is determined. Then, the corresponding target patient's skin wound grading result is output to provide a basis for subsequent diagnosis and treatment.

[0085] For example, in the clinical grading standard, the grading thresholds are set as follows: 1 to 2 corresponds to Grade I, which is well healed; 2 to 3 corresponds to Grade II, which is in the middle stage of healing; and 3 and above corresponds to Grade III, which requires intervention. Then, the comprehensive assessment value of 2.0 is compared with the clinical grading standard, and 2.0 corresponds to Grade II. The patient's skin wound grading result is output as being in the middle stage of healing. Finally, this result will be used by medical staff to formulate the patient's subsequent dressing change frequency and rehabilitation training plan, and will also be stored in the patient's electronic medical record to provide a reference for subsequent diagnosis and follow-up.

[0086] This application avoids the one-sidedness of single-feature assessment by combining three core features reflecting wound condition—area, texture, and color—with clinical standards for multi-dimensional comprehensive evaluation. Then, through quantification and weighted fusion of graded parameters, features from different dimensions are scientifically integrated, allowing the comprehensive assessment value to fully reflect the severity and healing status of the wound. Finally, the grading results, based on clinical thresholds and aligned with medical diagnostic and treatment needs, provide a reliable basis for medical staff to develop precise treatment and nursing plans, thereby effectively improving the scientific rigor and clinical applicability of wound assessment.

[0087] Figure 3 This is a schematic diagram illustrating a specific implementation of a deep learning-based skin wound area measurement and grading system provided in this application. (Refer to...) Figure 3 The system may include: The acquisition module 31 is used to acquire the initial visible light image and the initial near-infrared light image of the target patient's skin wound.

[0088] The fusion module 32 is used to perform image enhancement processing on the initial visible light image and the initial near-infrared light image respectively to obtain the target visible light image and the target near-infrared light image, and to fuse the target visible light image and the target near-infrared light image to obtain the skin wound image.

[0089] The segmentation module 33 is used to segment the skin wound image using a semantic segmentation algorithm to obtain a binarized image, and to calculate the number of pixels belonging to the skin wound in the binarized image to obtain the actual wound area.

[0090] The extraction module 34 is used to perform deep learning-based feature extraction processing on the skin wound image to obtain the texture roughness and color uniformity features of the skin wound.

[0091] The grading module 35 is used to grade the skin wound based on the actual wound area, the texture roughness, and the color uniformity characteristics, combined with a preset clinical skin wound grading standard, to obtain the skin wound grading result of the target patient.

[0092] The deep learning-based skin wound area measurement and grading system of this application is used to implement the aforementioned deep learning-based skin wound area measurement and grading method. Therefore, the specific implementation of the deep learning-based skin wound area measurement and grading system can be found in the embodiment section of the deep learning-based skin wound area measurement and grading method above. The specific implementation can be referred to the description of the corresponding embodiment, and will not be repeated here.

[0093] like Figure 4 As shown, this application also provides an electronic device, including: a memory 41 for storing a computer program; and a processor 42 for executing the computer program to implement the steps of any of the deep learning-based skin wound area measurement and grading methods described above.

[0094] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the above-described deep learning-based skin wound area measurement and grading methods.

[0095] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0096] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the embodiments of the deep learning-based skin wound area measurement and grading method described above.

[0097] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0098] The foregoing provides a detailed description of a deep learning-based method and system for measuring and grading skin wound area. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of these embodiments are merely illustrative of the method and its core concepts. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A deep learning-based method for measuring and grading skin wound area, characterized in that, include: Acquire initial visible light and initial near-infrared light images of the target patient's skin wound; Image enhancement processing is performed on the initial visible light image and the initial near-infrared light image respectively to obtain the target visible light image and the target near-infrared light image. The target visible light image and the target near-infrared light image are then fused to obtain the skin wound image. A semantic segmentation algorithm is used to segment the skin wound image to obtain a binarized image. The number of pixels belonging to the skin wound in the binarized image is calculated to obtain the actual wound area. The skin wound image is subjected to feature extraction processing based on deep learning to obtain the texture roughness and color uniformity features of the skin wound. Based on the actual wound area, the texture roughness, and the color uniformity, and combined with the preset clinical skin wound grading standards, the skin wound is graded to obtain the skin wound grading results for the target patient.

2. The method according to claim 1, characterized in that, The process involves employing a semantic segmentation algorithm to segment the skin wound image, obtaining a binarized image, and calculating the number of pixels belonging to the skin wound in the binarized image to determine the actual wound area. This includes: The skin wound image is processed to obtain multi-scale depth features. Based on the multi-scale depth features, a semantic segmentation algorithm is used to obtain the boundary contour between the wound area and the normal skin area in the skin wound image. Based on the boundary contour, the pixels in the skin wound image are divided into wound region pixels and background region pixels, and a preliminary segmentation image is generated based on the wound region pixels and the background region pixels; The preliminary segmented image is subjected to contour optimization processing to obtain a binarized image, and the total number of pixels in the wound area in the binarized image is counted. Based on the total number of pixels in the wound area and the preset actual size of the unit pixel, the actual wound area of ​​the skin wound is obtained.

3. The method according to claim 2, characterized in that, The skin wound image is processed to obtain multi-scale depth features. Based on these multi-scale depth features, a semantic segmentation algorithm is used to obtain the boundary contour between the wound area and the normal skin area in the skin wound image, including: Multiple processing layers with different receptive fields are used to extract layered features from the skin wound image, resulting in shallow features containing the fine structure of the wound and deep features containing the overall morphology of the wound. The shallow features and the deep features are fused to obtain multi-scale depth features, which are then used to construct a feature map. Based on the multi-scale depth features, a semantic segmentation algorithm with an edge-aware feature enhancement module is used to perform pixel-level classification of the skin wound image. During the pixel-level classification process, the edge-aware feature enhancement module calculates the gradient information of the feature map to enhance the feature response at the junction of wound tissue and normal skin, thereby obtaining the pixels of wound tissue and normal skin tissue. Based on the pixels of the wound tissue and the pixels of the normal skin tissue, the boundary point between the wound area and the normal skin area is obtained. Adjacent boundary points are connected to form the boundary contour between the wound area and the normal skin area.

4. The method according to claim 2, characterized in that, The process of performing contour optimization on the initially segmented image to obtain a binarized image, and counting the total number of pixels in the wound region of the binarized image, includes: Connectivity analysis is performed on the preliminary segmented image to obtain discontinuous regions between boundary points; Based on the near-infrared band information and visible light color information in the skin wound image, the tissue characteristics of the discontinuous region are determined. According to the tissue characteristics, an adaptive interpolation algorithm is used to complete the boundary in order to establish a complete closed boundary contour. The closed boundary contour is smoothed and filtered to obtain a binarized image; In the binarized image, the connected pixel groups of the wound area are identified, and the total number of pixels contained in the connected pixel groups is counted to obtain the total number of pixels in the wound area.

5. The method according to claim 1, characterized in that, The deep learning-based feature extraction process on the skin wound image to obtain the texture roughness and color uniformity features of the skin wound includes: Based on the binarized image, the wound area image is extracted from the skin wound image; The wound area image is input into a pre-trained feature extraction network, which includes parallel texture feature branches and color feature branches; In the texture feature branch, multiple Gabor filter layers and the first convolutional layer are used to extract deep texture features of the wound area. Based on the deep texture features of the wound area, the gray value variation amplitude between adjacent pixels in the wound area image is calculated to obtain the texture roughness. In the color feature branch, the color distribution is learned through the second convolutional layer, and based on the color distribution, the concentration of different color channel values ​​in the wound area image is statistically analyzed to obtain the color uniformity feature.

6. The method according to claim 1, characterized in that, The process involves grading the skin wound based on the actual wound area, texture roughness, and color uniformity, combined with a preset clinical skin wound grading standard, to obtain the skin wound grading result for the target patient, including: The actual wound area is matched with the area threshold in the preset clinical skin wound grading standard to obtain the area grading parameter. The texture roughness is compared with the texture reference feature in the preset clinical skin wound grading standard to obtain the texture grading parameter. The color uniformity feature is analyzed with the color distribution benchmark in the preset clinical skin wound grading standard to obtain the color grading parameter. The area grading parameters, texture grading parameters, and color grading parameters are input into a preset multi-dimensional grading decision model. The parameters are then weighted and fused through the preset multi-dimensional grading decision model to obtain a comprehensive evaluation value. The comprehensive assessment value is compared with the grading threshold in the preset clinical skin wound grading standard to obtain the skin wound grading result of the target patient.

7. The method according to claim 1, characterized in that, The process involves performing image enhancement processing on the initial visible light image and the initial near-infrared light image respectively to obtain a target visible light image and a target near-infrared light image, and then fusing the target visible light image and the target near-infrared light image to obtain a skin wound image, including: The initial visible light image and the initial near-infrared light image are subjected to orientation transformation and brightness adjustment operations, respectively, to obtain the target visible light image and the target near-infrared light image; Based on the target visible light image, the edge contour information of the skin wound is obtained; based on the target near-infrared light image, the tissue depth information of the skin wound is obtained. The edge contour information and the tissue depth information are combined to obtain a skin wound image.

8. A deep learning-based skin wound area measurement and grading system, characterized in that, include: The acquisition module is used to acquire initial visible light and initial near-infrared light images of the target patient's skin wound. The fusion module is used to perform image enhancement processing on the initial visible light image and the initial near-infrared light image respectively to obtain the target visible light image and the target near-infrared light image, and to fuse the target visible light image and the target near-infrared light image to obtain the skin wound image; The segmentation module is used to segment the skin wound image using a semantic segmentation algorithm to obtain a binarized image, and to calculate the number of pixels belonging to the skin wound in the binarized image to obtain the actual wound area. The extraction module is used to perform deep learning-based feature extraction processing on the skin wound image to obtain the texture roughness and color uniformity features of the skin wound. The grading module is used to grade the skin wound based on the actual wound area, the texture roughness, and the color uniformity, combined with a preset clinical skin wound grading standard, to obtain the skin wound grading result of the target patient.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the deep learning-based skin wound area measurement and grading method as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the implementation of the deep learning-based skin wound area measurement and grading method as described in any one of claims 1 to 7.

Images