A Pressure Injury Staging Recognition Method and Device Based on Multi-Feature Fusion

Abstract

The present invention relates to the technical field of pressure injury detection, and discloses a method and device for identifying the stage of pressure injury based on multi-feature fusion. The method includes: processing and analyzing a color image, a depth image, and a thermal imaging image of the pressure injury site to extract the following three features: a first attribute feature, a second attribute feature, and a third attribute feature. The first attribute feature includes the area, color, and surface flatness of the injury site. The second attribute feature includes the temperature difference and the temperature distribution feature. The third attribute feature includes the skin humidity information and the skin pH information. Inputting the first attribute feature, the second attribute feature, and the third attribute feature into a comprehensive evaluation model, and performing attribute feature fusion calculation by the comprehensive evaluation model to obtain the pressure injury grade of the injury site. The present invention can accurately evaluate and judge the grade of pressure injury, so as to establish a scientific and unified evaluation standard.

Description

Technical Field

[0001] The present invention relates to the technical field of pressure injury detection, and particularly to a method and device for stage recognition of pressure injury based on multi-feature fusion. Background Art

[0002] Pressure Injury (PI), also known as pressure ulcer, bedsore, is a kind of skin and subcutaneous tissue injury commonly found in the bony prominences of the human body. Its occurrence mechanism is mainly caused by the combined action of pressure or pressure combined with shear force. Accurate assessment of pressure injury is significantly helpful for its treatment effect.

[0003] Traditional manual clinical judgment mainly relies on the experience and intuition of medical staff, and has the following limitations:

[0004] 1. Strong subjectivity, different medical institutions or medical staff may adopt different assessment criteria;

[0005] 2. Lack of objective data support, which may lead to inaccurate judgment of the degree of skin injury, thus affecting the formulation of treatment plans, resulting in poor treatment effects or extended treatment cycles;

[0006] 3. The assessment process often requires the patient to maintain a specific posture or undergo multiple examinations, which may increase the discomfort of the patient.

[0007] Existing computer-aided assessment methods still have obvious deficiencies: the assessment indicators are single, lacking the collection and fusion analysis of multi-modal indicators, resulting in insufficient assessment accuracy. Summary of the Invention

[0008] In view of this, the present invention provides a method and device for stage recognition of pressure injury based on multi-feature fusion, which solves the problems that traditional manual clinical judgment depends on the experience and intuition of medical staff, has large subjectivity, and different medical institutions or medical staff may adopt different assessment criteria, lacks objective data support, and existing computer-aided pressure injury assessment methods have the problem of single assessment indicators, and lack of collection and fusion analysis of multiple groups of indicators resulting in insufficient accuracy. By combining technical means of multiple features, not only the accuracy and efficiency of stage recognition are improved, but also objective data support and standardized assessment methods are provided for clinical practice.

[0009] To achieve the above object, the present invention provides the following technical solutions:

[0010] A pressure injury staging recognition method based on multi - feature fusion, comprising: obtaining a color image, a depth image, and a thermal imaging image of the pressure injury site of the human body, preprocessing the color image and the depth image to obtain fused features, inputting the fused features into a trained first feature extraction model to obtain segmentation data of the injury site, and obtaining a first attribute feature T1 according to the segmentation data;

[0011] Performing grayscale processing on the thermal imaging image, performing normalization preprocessing on the pixel intensity after grayscale processing by using histogram equalization to enhance the gray - scale contrast between the injury area and normal tissues, inputting the preprocessed image data into a trained second feature extraction model to obtain a second attribute feature T2;

[0012] Measuring the humidity information and pH value information of the skin at the edge of the injury site through a humidity measuring element and a pH measuring element, and performing corresponding association between the measurement information and the edge position of the injury site in the color image to obtain a third attribute feature T3;

[0013] Inputting the first attribute feature T1, the second attribute feature T2, and the third attribute feature T3 into a comprehensive evaluation model, and performing attribute - feature fusion calculation by the comprehensive evaluation model to obtain the pressure injury grade of the injury site.

[0014] Further, the first attribute feature T1 includes the area, color, and surface flatness of the injury site; the second attribute feature T2 includes the temperature difference and temperature distribution attribute features; the third attribute feature T3 includes skin humidity information and skin pH value information.

[0015] Further, preprocessing the color image and the depth image to obtain fused features includes:

[0016] Performing bilateral filtering on the depth image by using a bilateral filtering algorithm, and performing coordinate alignment processing on the color image and the depth image to realize the mapping of color - image points to depth - image points, obtaining the color image and the depth image after coordinate alignment;

[0017] Performing edge detection on the aligned color image by using an edge - detection method based on the Canny operator, taking the outer contour of the obtained edge as the ROI region, and obtaining the ROI region of the aligned depth image based on the ROI region of the aligned color image and the mapping relationship between the depth image and the color image;

[0018] Converting the depth data corresponding to the depth image into three - dimensional point - cloud data to obtain three - dimensional point - cloud data with color information, registering the three - dimensional point - cloud data, and then reconstructing the human - body surface data by using a surface - reconstruction method of triangular meshing to obtain a colored human 3D model.

[0019] Further, the first feature extraction model includes a color space conversion module, a seed selection module, a segmentation threshold selection module, and a region merging module; the color space conversion module is used to convert the RGB color space to which the reconstructed color point cloud belongs into the HSV color space; the seed selection module is used to count the distribution of H values, take the local maximum points as seed points, and arrange them in order of priority to obtain a seed point queue until the first unmarked point dequeues, and then diffuse the seed points; the segmentation threshold selection module is used to calculate the optimal segmentation threshold of the seed points by using the adaptive threshold method, find the neighboring points of the seed points by using the neighboring point search method based on grid information and KNN statistical method, judge the similarity between the neighboring points and the seed points according to the optimal segmentation threshold, calculate the H values of the neighboring points, and if the H value is within the segmentation interval of the seed points, then perform similarity marking on the neighboring points, and iterate until all seed points are selected; the region merging module is used to merge adjacent regions according to the given color threshold, and merge the regions with a small number of point clouds into their neighboring regions according to the preset threshold to obtain the segmentation region image of the damaged part.

[0020] Furthermore, the first attribute feature T1 obtained according to the segmentation data includes:

[0021] According to the correspondence between the points in the region and the region of the segmentation region image, find all the points in the target region according to the user-marked points, find all the triangular meshes associated with the feature points in the region, remove the duplicate triangular meshes, and calculate the area of the sub-region by using the triangular patch fitting method. The vertices of a triangle are t1(x1, y1, z1), t2(x2, y2, z2), and t3(x3, y3, z3) respectively. Calculate the three side lengths a, b, and c of the triangle according to the coordinates, and calculate the area Si = , where s is the semi-perimeter, , and the total damaged area is , where n is the total number of triangular patches;

[0022] Obtain the color features of different damaged parts according to the segmentation data, calculate the proportion of each color pixel, compare the proportion with the preset threshold to divide the damage category, and use the number of the damage category as the color quantization value;

[0023] According to the wound depth information of the damaged part, obtain the point cloud with the maximum perpendicular distance from the skin fitting surface, and use this distance value as the flatness quantization value.

[0024] Further, the second feature extraction model includes a segmentation module and a temperature feature extraction module; the segmentation module is used to obtain a segmentation threshold and extract the range features of the damage area according to the gray threshold segmentation algorithm. The temperature feature extraction module is connected to the segmentation module and is used to obtain the temperature area where the temperature in the damage area is higher or lower than the preset body surface temperature threshold, calculate the difference e between the highest temperature and the lowest temperature, and the proportion information f of the area of the high-temperature area.

[0025] The segmentation module includes several steps of threshold calculation, region segmentation, and range extraction. Specifically, the optimal gray threshold T is iteratively calculated using the Otsu algorithm based on the between-class variance method, satisfying the minimization of the within-class variance between the damage area (high-gray area) and the background. The formula is: , where L is the maximum gray level, 、 is the proportion of two types of pixels, 、 is the class mean; when performing region segmentation and optimization, a binarization operation is performed to generate a mask image, isolated noise points are eliminated through morphological closing operations (dilation first and then erosion), and the continuous closed damage contour is extracted using the Canny edge detection method. The temperature feature extraction module includes steps of temperature data processing, high-temperature and low-temperature region determination, temperature difference calculation, and area statistics. Specifically, the coordinates of the segmented damage area are mapped to the original thermal image temperature matrix, and environmental radiation interference is eliminated based on the environmental temperature compensation algorithm. The formula is: , where, is the target true temperature, is the object surface radiation temperature directly measured by the infrared thermal imager, is the emissivity of the object being measured, and for human skin ≈0.95 0.98, is the environmental temperature. If =0.95, and the measured =38°C, and the environmental temperature =25°C, then the true temperature is: °C; when determining the high-temperature region, a dynamic body surface temperature threshold Tbase is set, and Tbase can be dynamically adjusted based on clinical data. The pixel set satisfying Tp≥Tbase+ΔT1 is extracted, where ΔTn is the pathological feature temperature difference constant. During region determination, the pixel set satisfying Tp≤Tbase+ΔT2 is extracted; when calculating the temperature difference feature parameters, the formula is used to calculate the maximum temperature difference value to characterize the temperature distribution dispersion in the damage area; the proportion f of the high-temperature area is obtained by counting the number of high-temperature pixels using the connected component labeling algorithm, and calculating its ratio to the total damage area , , and then output the result to the grading module for quantitatively evaluating the degree of tissue inflammation or the risk level of necrosis. Multidimensional diagnostic indicators are formed through quantitative evaluation. The damage contour and the pseudo-color mapping of the high-temperature area can be superimposed and displayed on the visual interaction interface.

[0026] Furthermore, the comprehensive evaluation model performs attribute feature fusion calculation to obtain the pressure injury level of the damaged part, including:

[0027] Obtain the weight coefficients wi corresponding to the preset first attribute feature T1, second attribute feature T2, and third attribute feature T3;

[0028] According to the weight coefficient wi, calculate the pressure injury evaluation value of the damaged part + , , , , is the corresponding grade value, is the quantified value of the area of the damaged part, is the quantified value of the color of the damaged part, is the quantified value of the surface flatness of the damaged part, is the quantified value of the temperature difference of the damaged part, is the quantified value of the temperature distribution attribute feature of the damaged part, is the quantified value of the skin humidity information of the damaged part, is the quantified value of the skin pH information of the damaged part, is the x-th quantified value of the i-th attribute feature.

[0029] The present invention also provides a pressure injury staging recognition device based on multi-feature fusion, which is implemented by using the above-mentioned pressure injury staging recognition method based on multi-feature fusion, and includes: an image acquisition module, a feature extraction module, and a grading module;

[0030] The image acquisition module is used to acquire the color image, depth image, and thermal imaging image of the pressure injury part of the human body, preprocess the color image and depth image to obtain fusion features, and is also used to measure the humidity information and pH value information of the skin at the edge of the injury part through a humidity measurement element and a pH measurement element, and correspond and associate the measurement information with the edge position of the injury part in the color image;

[0031] The feature extraction module is used to input the fused features into the trained first feature extraction model to obtain the segmentation data of the damaged part, obtain the first attribute feature T1 according to the segmentation data, perform grayscale processing on the thermal imaging image, perform normalization preprocessing on the pixel intensity after grayscale processing by using histogram equalization, input the preprocessed image data into the trained second feature extraction model to obtain the second attribute feature T2, quantify the measured humidity information and pH value information, and perform corresponding association between the measurement information and the edge position of the damaged part in the color image to obtain the third attribute feature T3;

[0032] The grade classification module is used to input the first attribute feature T1, the second attribute feature T2 and the third attribute feature T3 into the comprehensive evaluation model, and the comprehensive evaluation model performs attribute feature fusion calculation to obtain the pressure injury grade of the damaged part.

[0033] Furthermore, the grade classification module includes a weight coefficient acquisition module and a grade calculation module;

[0034] The weight coefficient acquisition module is used to obtain the weight coefficients wi of the first attribute feature T1, the second attribute feature T2 and the third attribute feature T3;

[0035] According to the weight coefficient wi, calculate the pressure injury evaluation value of the damaged part + , , , , is the corresponding grade value, is the x-th quantization value of the i-th attribute feature.

[0036] As can be seen from the above technical solutions, the advantages of the present invention are:

[0037] The present invention collects color and depth images of the patient's body, thermal imaging information of the pressure injury site, and information about the humidity and pH value of the skin at the edge of the injury site. The color and depth images are aligned and converted into three-dimensional color point cloud data to construct a color 3D human body model. An intelligent segmentation method based on color space is used to accurately identify and extract segmented regional images of the injury site. Combining color features with wound depth information, a quantization algorithm is used to accurately calculate quantitative indicators of the injury site. Temperature characteristics of the patient's injury site, as well as information about the humidity and pH value of the skin at the edge of the injury site, are also extracted. Through a multidimensional data fusion processing method, the pressure injury grade of the injury site is determined. This makes the pressure injury grade assessment more accurate and can establish scientific and unified assessment standards. This provides objective data support and standardized assessment methods for clinical practice, facilitating consensus on the occurrence of pressure injuries among different departments and institutions. Compared with traditional manual assessment methods, this method reduces the inconvenience of patients having to maintain specific postures or undergo multiple examinations. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0039] Figure 1 Schematic diagram of the steps of the pressure injury staging identification method based on multi-feature fusion of the present invention.

[0040] Figure 2 Schematic diagram of the structure of the pressure injury staging identification device based on multi-feature fusion of the present invention.

[0041] Figure 3 This is a schematic structural diagram of the device for collecting and analyzing pressure injury images of the present invention.

[0042] Figure 4 FIG. 4 is a schematic diagram of step S4 of this embodiment.

[0043] Figure 5 This is a configuration example diagram of the image acquisition device and the image acquisition and analysis device of this embodiment.

[0044] Figure 6 Schematic diagram of the pressure injury stages in this embodiment.

[0045] Reference numerals:

[0046] 10. Image acquisition module; 11. Feature extraction module; 12. Grade classification module; 121. Weight coefficient acquisition module; 122. Grade calculation module; 20. Processor; 30. Acquisition device. DETAILED DESCRIPTION

[0047] To make the objectives, technical solutions and advantages of the present invention more clear and understandable, the present invention will be further described in detail below in conjunction with the embodiments and the accompanying drawings. Herein, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but do not limit the present invention.

[0048] Referring to Figures 1 to 6 , this embodiment provides a pressure injury staging recognition method and device based on multi-feature fusion. The pressure injury staging recognition method based on multi-feature fusion can accurately evaluate and judge the level of pressure injury and establish a unified evaluation standard, specifically including:

[0049] Step S1: Obtain a color image, a depth image, and a thermal imaging image of the pressure injury site of the human body, preprocess the color image and the depth image to obtain fusion features, input the fusion features into a trained first feature extraction model to obtain segmentation data of the injury site, and obtain a first attribute feature T1 according to the segmentation data, where the first attribute feature T1 includes the area, color, and surface flatness of the injury site.

[0050] In this embodiment, by collecting a color image, a depth image, and a thermal imaging image of the patient's body, extracting image features, and performing fusion analysis on the image features, the level of skin pressure injury can be accurately evaluated. Among them, when collecting the color image and the depth image, a Kinect, a somatosensory peripheral device developed by Microsoft, is used. The structured light emitted by the Kinect will not cause harm to the human body, and it is small in size and convenient to scan the patient from multiple angles. It has three lenses. The lenses on the left and right sides are an infrared emitter and an infrared CMOS camera respectively. The infrared emitter and the infrared CMOS camera form a depth sensor, and a depth image can be generated through the emission and reception of infrared light. The middle lens is an RGB camera, and the RGB camera is used to obtain a color image.

[0051] Preprocess the color image and the depth image to obtain fused features, including: applying the bilateral filtering algorithm to the depth image for bilateral filtering, and performing coordinate alignment processing on the color image and the depth image to achieve the mapping from color image points to depth image points, obtaining the color image and the depth image after coordinate alignment. Since the depth image directly obtained by Kinect will have hole points at the edges or in occluded areas, resulting in poor image quality, it is necessary to filter the collected depth image. Bilateral filtering can maintain clear boundaries. The RGB camera and the depth sensor for collecting images are not in the same position, so it is necessary to align the color image and the depth image. Use the edge detection method based on the Canny operator to detect the edges of the aligned color image, and take the outer contour of the obtained edges as the ROI region. Based on the ROI region of the aligned color image and the mapping relationship between the depth image and the color image, obtain the ROI region of the aligned depth image; convert the depth data corresponding to the depth image into three-dimensional point cloud data, obtain three-dimensional point cloud data with color information, register the three-dimensional point cloud data, and then reconstruct the human body surface data through the surface reconstruction method of triangular meshing to obtain a colored human 3D model. Extract the human body contour data through edge detection, and perform further conversion and processing of the point cloud data based on the human body contour data. Among them, preprocessing such as point cloud registration can integrate and filter the point cloud through matching algorithms and filtering algorithms.

[0052] Pressure injury is caused by long-term pressure on local tissues, resulting in persistent ischemia, hypoxia, and malnutrition, and ultimately leading to tissue ulceration and necrosis. The prone sites of pressure sores mainly include bony prominences: the sacrococcygeal region, ischial tuberosities, scapulae, elbows, and heels, as well as the auricle, acromion, elbow, hip, inner and outer sides of the knee joint, medial and lateral malleoli, back, thigh, etc. Therefore, when calculating the area of the damaged site, the calculation method based on the 3D human model is more accurate than only identifying and calculating the damaged site based on a two-dimensional image. In this embodiment, since the depth image collected by Kinect is a two-dimensional depth image, it needs to be converted into three-dimensional point cloud data with three-dimensional coordinate information. To facilitate finding a specific damaged area, it is also necessary to match and overlay the color two-dimensional image and the converted point cloud data to obtain three-dimensional point cloud with color information. The three-dimensional point cloud with color information is obtained by processing a frame of depth image and a frame of RGB image obtained at the same time through the above process. After point cloud registration of the colored three-dimensional point cloud data, use the surface reconstruction method of triangular meshing to reconstruct the human body surface data to obtain a colored human 3D model, and then segment and extract features of the damaged site.

[0053] In this embodiment, the first feature extraction model is used to segment the acquired three-dimensional point cloud, and area, color, and surface flatness are extracted based on the segmented data. Specifically, the first feature extraction model includes a color space conversion module, a seed selection module, a segmentation threshold selection module, and a region merging module; the color space conversion module is used to convert the RGB color space to which the reconstructed color point cloud belongs into the HSV color space; the seed selection module is used to count the distribution of H values, take the local maximum points as seed points, and arrange them in order of priority to obtain a seed point queue until the first unmarked point leaves the queue, and this point is used as a seed point for diffusion; the segmentation threshold selection module is used to calculate the optimal segmentation threshold of the seed point by using the adaptive threshold method, find the neighboring points of the seed point by using the neighboring point search method based on grid information and KNN statistical method, judge the similarity between the neighboring points and the seed point according to the optimal segmentation threshold, calculate the H value of the neighboring points, and when the H value is within the segmentation interval of the seed point, perform similarity marking on the neighboring points, and iterate until all seed points are selected; the region merging module is used to merge adjacent regions according to the given color threshold, and merge regions with a small number of point clouds into their neighboring regions according to the preset threshold to obtain a segmentation region image of the damaged part.

[0054] In this embodiment, the RGB space is mapped to the HSV space through a non-linear transformation, an H value histogram is constructed, after smoothing the curve with a Savitzky-Golay filter, the local maximum points are detected by the first-order difference method, and the identified local maximum points are used as candidate seeds, which are arranged in descending order of peak height to form a priority queue. At initialization, the H value corresponding to the highest peak is selected as the first seed point. During the region growing process, when the region corresponding to the seed at the head of the queue is fully diffused, the activation of the next local maximum point is automatically triggered until all candidate peaks are traversed or the termination condition is met (such as covering 95% of the point cloud). The dynamic octree grid division is used to accelerate the neighboring point search. The point cloud space is divided into cubic grids with side length l = 2r. The KNN search is performed on the grid where the seed point is located and its adjacent grids to ensure that all points within the radius r are covered. The dynamic threshold TH is calculated based on the neighborhood statistics of the seed point. , where is the mean value of the neighborhood ΔH, and σΔH is the standard deviation. Points satisfying ΔH ≤ TH are marked as similar points and added to the current region, and the index of the processed point cloud is recorded to prevent repeated calculation. When merging adjacent regions according to the given color threshold, calculate the HSV mean vector of the adjacent regions , if is satisfied, then region fusion is performed, where , is the Euclidean distance of the color difference, is the norm of the region color intensity. Then set the threshold , for the number of points N < For isolated areas, calculate the distance from their centroid to all adjacent areas and merge them into the adjacent large area with the closest space and the smallest color difference.

[0055] The HSV color space decomposes color information into hue (H), saturation (S), and value (V). Data processing using the HSV color space is superior to that of the RGB color space, enabling better segmentation of point clouds. To reduce the impact of illumination, this embodiment uses hue (H) as the similarity metric for point clouds, and segmentation is performed based on a region-growing algorithm. Region-growing requires finding the points with the strongest correlation for diffusion. Because the distance between two points in a three-dimensional point cloud is not necessarily the true surface distance, this embodiment uses grid information and a KNN statistical method to find neighboring points. Grid information, as correlation information, can avoid errors in point segmentation.

[0056] The first attribute feature T1 obtained according to the segmented data includes:

[0057] According to the correspondence between points and regions in the segmented region image (because the patient's body may include multiple different injury areas, when processing, the area to which the point belongs can be quickly found based on the association information of the above-mentioned points and their respective areas), all points in the target area are found according to the user's marked points (after determining the target area, all points in the area can be quickly found based on the information of the regional points in the above-mentioned region growth process), all triangular meshes associated with the feature points are found in the area, duplicate triangular meshes are removed, and the area of the sub-region is calculated using the triangular patch fitting method. The vertices of a triangle are t1(x1, y1, z1), t2(x2, y2, z2), and t3(x3, y3, z3). The three side lengths a, b, and c of the triangle are calculated according to the coordinates, and the area Si= is calculated according to Heron's formula. , where s is the semi-perimeter, The total damaged area is , where n is the total number of triangles;

[0058] Based on the segmented data, the color features of different damaged areas are obtained. The proportion of pixels of each color is calculated based on the RGB color features. The proportion is compared with a preset threshold to classify the damage category, and the damage category number is used as the color quantization value. Based on the wound depth information of the damaged area, the point cloud with the maximum vertical distance from the skin fitting surface is obtained, and this distance value is used as the flatness quantization value. Because this embodiment uses a triangulated mesh surface reconstruction method to reconstruct human surface data, when calculating the damaged area, the mesh information in the reconstruction stage can be used to count the corresponding triangular facets to estimate the damaged area.

[0059] Step S2: Grayscale the thermal imaging image, perform normalization preprocessing on the pixel intensity after grayscale processing using histogram equalization to enhance the grayscale contrast between the damaged area and normal tissues, and input the preprocessed image data into the trained second feature extraction model to obtain the second attribute feature T2. The second attribute feature T2 includes the temperature difference and the temperature distribution attribute feature. The thermal imaging image is obtained by a thermal imager.

[0060] The second feature extraction model includes a segmentation module and a temperature feature extraction module; the segmentation module is used to obtain a segmentation threshold and extract the range feature of the damaged area according to the grayscale threshold segmentation algorithm. The temperature feature extraction module is connected to the segmentation module and is used to obtain the temperature area where the temperature in the damaged area is higher or lower than the preset body surface temperature threshold, calculate the difference e between the highest temperature and the lowest temperature, and the proportion information f of the area of the high-temperature area.

[0061] In this embodiment, the segmentation module includes several steps of threshold calculation, region segmentation, and range extraction. Specifically, the optimal grayscale threshold T is iteratively calculated using the Otsu algorithm (maximum between-class variance method) to minimize the within-class variance between the damaged area (high grayscale area) and the background. The formula is: , where L is the maximum gray level, , is the proportion of two types of pixels, , are the class means; when performing region segmentation and optimization, a binary operation is performed to generate a mask image, isolated noise points are eliminated through morphological closing operations (first dilation and then erosion), and the continuous closed damage contour is extracted using the Canny edge detection method. The temperature feature extraction module includes steps of temperature data processing, determination of high-temperature and low-temperature regions, temperature difference calculation, and area statistics. Specifically, the coordinates of the segmented damaged area are mapped to the original thermal image temperature matrix, and environmental radiation interference is eliminated based on the environmental temperature compensation algorithm. The formula is: , where, is the target true temperature, is the surface radiation temperature of the object directly measured by the infrared thermal imager, is the emissivity of the object being measured. For human skin, ≈0.95 0.98, is the environmental temperature. If =0.95, and the measured =38°C, and the environmental temperature =25°C, then the true temperature is: °C; When determining the high - temperature area, a dynamic body surface temperature threshold Tbase is set. Tbase can be dynamically adjusted based on clinical data. A pixel set satisfying Tp≥Tbase + ΔT1 is extracted, where ΔTn is a pathological feature temperature difference constant. When performing area determination, a pixel set satisfying Tp≤Tbase + ΔT2 is extracted; when calculating the temperature difference feature parameters, the formula is used to calculate the maximum temperature difference value to characterize the temperature distribution dispersion of the damaged area; the high - temperature area ratio f is obtained by counting the number of high - temperature pixels through the connected - component labeling algorithm , calculating its ratio to the total damaged area , , and then the result is output to the grading module for quantitatively evaluating the degree of tissue inflammation or the risk level of necrosis. A multi - dimensional diagnostic index is formed through quantitative evaluation. The damage contour and the pseudo - color mapping of the high - temperature area can be superimposed and displayed on the visual interaction interface.

[0062] Step S3: Measure the humidity information and pH value information of the skin at the edge of the damaged area through a humidity measurement element and a pH measurement element, and make a corresponding association between the measurement information and the edge position of the damaged area in the color image to obtain the third attribute feature T3. The third attribute feature T3 includes skin humidity information and skin pH information.

[0063] In this embodiment, the humidity measurement element uses a polymer - film humidity - sensitive capacitance sensor, which can detect the change of moisture information on the skin by contacting the skin, convert the output capacitance signal into a pulse signal through a pre - processing circuit for analysis. In this embodiment, the method of taking the average value of multiple measurements is used to obtain the skin humidity measurement value, and the edge position of the damaged area is marked.

[0064] The pH measurement element uses a ROSS - series pH electrode, which is a planar probe and can measure the pH value of the solid or semi - solid surface. The signal collected by the pH electrode can be processed into a digital signal through an amplifier, an A / D converter, etc. for analysis. In this embodiment, the method of taking the average value of multiple measurements is used to obtain the measurement value of the skin pH value, and the edge position of the damaged area is marked.

[0065] Step S4: Input the first attribute feature T1, the second attribute feature T2, and the third attribute feature T3 into the comprehensive evaluation model, and the comprehensive evaluation model performs attribute - feature fusion calculation to obtain the pressure - injury grade of the damaged area.

[0066] Specifically, as Figure 4 shown, the comprehensive evaluation model performs attribute - feature fusion calculation to obtain the pressure - injury grade of the damaged area, including:

[0067] Step S41: Obtain the weight coefficients wi of the preset first attribute feature T1, second attribute feature T2, and third attribute feature T3. The weight coefficients wi can be obtained through learning by a machine learning model. In this embodiment, the decision tree method is adopted.

[0068] Step S42: Calculate the pressure injury assessment value of the damaged part according to the weight coefficient wi + , , , , is the corresponding grade value, is the quantified value of the area of the damaged part, is the quantified value of the color of the damaged part, is the quantified value of the surface flatness of the damaged part, is the quantified value of the temperature difference of the damaged part, is the quantified value of the temperature distribution attribute feature of the damaged part, is the quantified value of the skin humidity information of the damaged part, is the quantified value of the skin pH information of the damaged part, is the x-th quantified value of the i-th attribute feature.

[0069] In this embodiment, pressure injuries are divided into four stages (the results of pressure injuries in each stage are divided into three grades according to the size of the damaged area, that is, the pressure injury grades altogether include 12 grades. The damaged area less than 5% of the total body surface area belongs to grade 1, the damaged area greater than 5% and less than 10% of the total body surface area belongs to grade 2, and the damaged area greater than 10% of the total body surface area belongs to grade 3). In stage I pressure injuries, the manifestations are as follows: Color: Erythema (dark skin may be purple or dark red, red pigment is greater than 90%, and no black or yellow pigment is seen), belonging to category 1; Surface flatness: The skin is intact, the surface is flat, without damage, and the flatness quantified value is 0. Temperature: Slightly higher or lower than the surrounding skin (local inflammation or ischemia), the humidity value is usually normal, and the temperature difference between slightly moist (such as sweating) and normal body surface temperature is within ±2°C; pH value: Close to normal skin (about 5.4 - 5.9).

[0070] In stage II pressure injury, the manifestations are as follows: Color: The wound surface is pink or red (there may be serous blisters, with red pigment accounting for 75%-90%, and no black or yellow pigment is seen), belonging to category 2; Surface flatness: The epidermis and part of the dermis are missing, and blisters or superficial ulcers may form. The epidermis is damaged, the surface is moist or crusted, and the edges are clear. The flatness quantification value is 0 mm - 1.5 mm; Temperature: The temperature of the wound surface increases (inflammatory reaction), and the temperature difference from the normal body surface temperature is less than +3°C. Humidity value: Significantly moist (increased exudate). pH value: May increase to 6.0 - 7.5 (the exudate is slightly alkaline).

[0071] In stage III pressure injury, the manifestations are as follows: Color: Yellow necrotic tissue, red granulation, or black eschar can be seen on the wound surface (red pigment is less than 75%-95%, and black and yellow pigments appear), belonging to category 3; Surface flatness: The full-thickness skin is missing, involving the subcutaneous tissue but not reaching the fascia layer. There is an obvious depression, and the flatness quantification value is -1.5 mm - 0 mm; Temperature: The temperature of the necrotic area is low, and the surrounding tissues may have an increased temperature. The center is hypothermic (ischemic necrosis), and the periphery is hyperthermic; Humidity value: Highly moist (a large amount of exudate or necrosis liquefaction); pH value: Further increases to 7.0 - 8.5 (alkaline environment caused by the decomposition of necrotic tissue).

[0072] In stage IV pressure injury, the manifestations are as follows: Color: Black eschar (necrosis), yellow slough, or red granulation tissue is formed (red pigment is less than 75%-95%, and black and yellow pigments appear), belonging to category 3; Surface flatness: The full-thickness tissue is missing, exposing muscle, bone, or tendon. There is a deep depression, and the flatness quantification value is less than -1.5 mm; Temperature: The temperature of the necrotic area is low, the center is hypothermic (ischemic necrosis), and the surrounding area has a fever due to infection; Humidity value: Extremely moist (purulent exudate or slough liquefaction); pH value: Exceeds 8.0 (severe infection or alkalization of chronic wounds).

[0073] As Figure 6 shown, they are respectively the detection schematic diagrams of stage I grade I, stage II grade I, stage III grade I, and stage IV grade I pressure injuries.

[0074] According to the data shown in the following comparative examples, compared with the traditional manual clinical judgment method and the single-modal pressure injury assessment method based on computer assistance, the present invention is more objective and accurate. The experimental data set includes 200 pressure injury samples (including stages I - IV). The 200 pressure injury samples include 152 clinical case images from the wound repair department of a certain tertiary hospital and 48 images from the public medical imaging database. All samples have been verified for staging labels through pathological biopsy or CT imaging.

[0075] Comparative Example 1: The traditional visual assessment method was adopted (collecting independent assessment data of 2 senior wound therapists); Comparative Example 2: The single-modal analysis method was adopted (only using the color attribute features extracted from color images); Comparative Example 3: The multi-modal comprehensive evaluation model was adopted (using the complete three-modal feature fusion method). It can be seen from the experimental results that the accuracy rate of Comparative Example 1 was 68.2%, the sensitivity (stages I-II) was 65.4%, the specificity (stages III-IV) was 71.8%, and the AUC value was 0.72; the accuracy rate of Comparative Example 2 was 72.5% (95%CI 67.8-76.9), the sensitivity (stages I-II) was 73.1%, the specificity (stages III-IV) was 75.2%, and the AUC value was 0.76. The accuracy rate of Comparative Example 3 was 89.6% (95%CI 86.2-92.4), the Kappa value reached 0.86 (p<0.001), the sensitivity (stages I-II) was 92.3%, the specificity (stages III-IV) was 90.7%, and the AUC value was 0.94.

[0076] The clinical validation test carried out by a certain tertiary hospital (n = 152 patients, a total of 243 injury sites) showed that the multi-modal comprehensive evaluation model had a high consistency with the results of histopathological examination. Among them, the recognition sensitivity of stage I was 92.3%, and the area measurement error of stages III / IV was <5% (compared with the CT measurement results). Through cross-validation (5-fold CV), it was confirmed that the AUC value of the multi-modal comprehensive evaluation model reached 0.94 (95%CI 0.91-0.97), which was significantly better than the single imaging examination method. The clinical follow-up data showed that using this model to guide the nursing plan could reduce the deterioration rate of injuries above stage II by 41%, fully verifying the clinical value of multi-modal feature fusion in the assessment of pressure injuries.

[0077] In another embodiment, as Figure 5 shown, one or more of the acquisition devices 30 can also be integrally arranged for convenient detection.

[0078] As Figure 2 shown, this embodiment also discloses a pressure injury staging recognition device based on multi-feature fusion, which is implemented by using the above-mentioned pressure injury staging recognition method based on multi-feature fusion, and includes: an image acquisition module, a feature extraction module, and a grade division module;

[0079] The image acquisition module is used to acquire the color image, depth image, and thermal imaging image of the pressure injury site of the human body, preprocess the color image and depth image to obtain fusion features, and is also used to measure the humidity information and pH value information of the skin at the edge of the injury site through the humidity measurement element and pH measurement element, and perform corresponding association between the measurement information and the edge position of the injury site in the color image;

[0080] The feature extraction module is used to input the fused features into the trained first feature extraction model to obtain the segmentation data of the damaged part, obtain the first attribute feature T1 according to the segmentation data, perform grayscale processing on the thermal imaging image, perform normalization preprocessing on the pixel intensity after grayscale processing by using histogram equalization, input the preprocessed image data into the trained second feature extraction model to obtain the second attribute feature T2, quantify the measured humidity information and pH value information, and perform corresponding association between the measured information and the edge position of the damaged part in the color image to obtain the third attribute feature T3;

[0081] The grade classification module is used to input the first attribute feature T1, the second attribute feature T2, and the third attribute feature T3 into the comprehensive evaluation model, and the comprehensive evaluation model performs attribute feature fusion calculation to obtain the pressure injury grade of the damaged part. The grade classification module includes a weight coefficient acquisition module and a grade calculation module; the weight coefficient acquisition module is used to obtain the weight coefficients wi of the first attribute feature T1, the second attribute feature T2, and the third attribute feature T3; according to the weight coefficients wi, calculate the pressure injury evaluation value of the damaged part + , , , , is the corresponding grade value, is the x-th quantization value of the i-th attribute feature.

[0082] As Figure 3 shown, a device for collecting and analyzing pressure injury images includes:

[0083] A storage medium with computer program instructions stored thereon;

[0084] At least one processor 20, configured to execute the computer program instructions in the storage medium to implement the above-mentioned method for staging and identifying pressure injuries based on multi-feature fusion.

[0085] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, various changes and modifications can be made to the embodiments of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A pressure injury staging recognition device based on multi-feature fusion, characterized in that, Including: An image acquisition module, a feature extraction module, and a grading module; The image acquisition module is used to acquire a color image, a depth image, and a thermal imaging image of the pressure injury site of the human body, preprocess the color image and the depth image to obtain fusion features, and is also used to measure the humidity information and pH value information of the skin at the edge of the injury site through a humidity measurement element and a pH measurement element, and associate the measurement information with the edge position of the injury site in the color image; The feature extraction module is used to input the fusion features into a trained first feature extraction model to obtain segmentation data of the injury site, obtain a first attribute feature T1 according to the segmentation data, perform grayscale processing on the thermal imaging image, perform normalization preprocessing on the pixel intensity after grayscale processing by using histogram equalization, input the preprocessed image data into a trained second feature extraction model to obtain a second attribute feature T2, quantify the measured humidity information and pH value information, and associate the measurement information with the edge position of the injury site in the color image to obtain a third attribute feature T3; The grading module is used to input the first attribute feature T1, the second attribute feature T2, and the third attribute feature T3 into a comprehensive evaluation model, and perform attribute feature fusion calculation by the comprehensive evaluation model to obtain the pressure injury grade of the injury site; Preprocessing the color image and the depth image to obtain fusion features includes: Performing bilateral filtering on the depth image by using a bilateral filtering algorithm, and performing coordinate alignment processing on the color image and the depth image to realize the mapping of color image points to depth image points, so as to obtain the color image and the depth image after coordinate alignment; Performing edge detection on the aligned color image by using an edge detection method based on the Canny operator, taking the outer contour of the obtained edge as the ROI region, and obtaining the ROI region of the aligned depth image based on the ROI region of the aligned color image and the mapping relationship between the depth image and the color image; Converting the depth data corresponding to the depth image into three-dimensional point cloud data to obtain three-dimensional point cloud data with color information, registering the three-dimensional point cloud data, and then reconstructing the human body surface data by using a surface reconstruction method of triangular meshing to obtain a colored human body 3D model; The first feature extraction model includes a color space conversion module, a seed selection module, a segmentation threshold selection module, and a region merging module; the color space conversion module is used to convert the RGB color space to which the reconstructed color point cloud belongs into the HSV color space; the seed selection module is used to count the distribution of H values, take the local maximum points as seed points, and arrange them in order of priority to obtain a seed point queue until the first unmarked point dequeues, and diffuse the seed points; the segmentation threshold selection module is used to calculate the optimal segmentation threshold of the seed points by using the adaptive threshold method, find the neighboring points of the seed points by using the neighboring point search method based on grid information and KNN statistical method, judge the similarity between the neighboring points and the seed points according to the optimal segmentation threshold, calculate the H value of the neighboring points, and when the H value is within the segmentation interval of the seed points, perform similarity marking on the neighboring points, and iterate until all the seed points are selected; the region merging module is used to merge adjacent regions according to the given color threshold, and merge the regions with a small number of point clouds into their neighboring regions according to the preset threshold to obtain the segmentation region image of the damaged part; The first attribute feature T1 obtained according to the segmentation data includes: According to the correspondence between the in-region points and the regions of the segmented region images, find all the points within the target region based on the user-marked points, find all the triangular meshes associated with the feature points within the region, remove the duplicate triangular meshes, and calculate the area of the sub-region using the triangular patch fitting method. The vertices of a triangle are t1(x1, y1, z1), t2(x2, y2, z2), and t3(x3, y3, z3) respectively. Calculate the three side lengths a, b, and c of the triangle according to the coordinates, and calculate the area Si according to Heron's formula , where s is the semi-perimeter, , and the total damaged area is , where n is the total number of triangular patches; Obtain the color features of different damaged parts according to the segmentation data, calculate the proportion of each color pixel according to the color features, compare the proportion with the preset threshold to divide the damage categories, and use the number of the damage category as the color quantization value. According to the wound depth information of the damaged part, obtain the point cloud with the largest vertical distance from the skin fitting surface, and use this distance value as the flatness quantization value.

2. The pressure injury staging recognition device based on multi-feature fusion according to claim 1, wherein The first attribute feature T1 includes the area, color, and surface flatness of the damaged part; the second attribute feature T2 includes the temperature difference and temperature distribution attribute features; the third attribute feature T3 includes the skin humidity information and skin pH information.

3. The pressure injury staging recognition device based on multi-feature fusion according to claim 1, characterized in that, The second feature extraction model includes a segmentation module and a temperature feature extraction module; the segmentation module is used to obtain the segmentation threshold and implement the range feature extraction of the damaged area according to the gray threshold segmentation algorithm, the temperature feature extraction module is connected to the segmentation module, and the temperature feature extraction module is used to obtain the temperature area in the damaged area where the temperature is higher or lower than the preset body surface temperature threshold, calculate the difference e between the highest temperature and the lowest temperature, and the proportion information f of the area of the high-temperature area.

4. The pressure injury staging recognition device based on multi-feature fusion according to claim 1, characterized in that, The comprehensive evaluation model performs attribute feature fusion calculation to obtain the pressure injury grade of the damaged part, including: Obtain the weight coefficients wi corresponding to the preset first attribute feature T1, the second attribute feature T2, and the third attribute feature T3; Calculate the pressure injury evaluation value b of the damaged part according to the weight coefficient wi, + , , , , is the corresponding level value, is the quantification value of the area of the damaged part, is the quantification value of the color of the damaged part, is the quantification value of the surface flatness of the damaged part, is the quantification value of the temperature difference of the damaged part, is the quantification value of the temperature distribution attribute feature of the damaged part, is the quantification value of the skin humidity information of the damaged part, is the quantification value of the skin pH information of the damaged part.

Images