A wound image analysis method for assisting in diabetic foot ulcer staging

Abstract

The application discloses a kind of auxiliary diabetic foot ulcer grading wound image analysis method, it is related to medical image analysis technical field, can solve how to based on single frame optical image automatically, accurately distinguish the problem of solid-state necrotic tissue and liquid pus exudate with visual feature highly similar in diabetic foot ulcer wound, including: determining ulcer lesion area in input image, and determine lesion equivalent diameter;Determine the first point set that all tissue sample points to be classified in ulcer lesion area constitute, and the second point set that all potential convergence positions in ulcer lesion area constitute;According to lesion equivalent diameter, first point set and second point set, construct total cost matrix;Total cost matrix is solved using preset matching algorithm, each point in first point set is classified as liquid attribute point or solid attribute point;According to the number of being classified as liquid attribute point and first quantity determine liquid component proportion, compare with preset threshold value, output diabetic foot ulcer grading result.

Description

Technical Field

[0001] This invention relates to the field of medical image analysis technology, specifically to a wound image analysis method for assisting in the grading of diabetic foot ulcers. Background Technology

[0002] Diabetic foot ulcers are a common and serious complication in diabetic patients. Clinical severity grading (such as the Wagner classification system) is crucial for developing treatment plans including debridement and anti-infection therapy. Accurately distinguishing between deep ulcers and deep abscesses is particularly important. In clinical practice, current technologies often employ digital image analysis to assist in wound assessment. These methods typically rely on extracting and statistically analyzing visual features such as color and texture from wound images.

[0003] However, diabetic foot ulcers often contain both yellowish-white solid necrotic tissue (sludge) and liquid exudate (pus). The two exhibit a high degree of color and texture similarity in ordinary optical images. This "metachromatic" phenomenon makes it difficult to distinguish the physical states of the two in a stable and accurate manner based on traditional image feature analysis methods. As a result, the reliability of automated or remote assessment systems is insufficient, which can easily lead to misjudgment of the depth of infection. Summary of the Invention

[0004] To address the current technical problem of how to automatically and accurately distinguish between solid necrotic tissue and liquid purulent exudate with highly similar visual features in diabetic foot ulcer wounds based on single-frame optical images, the present invention aims to provide a wound image analysis method to assist in the grading of diabetic foot ulcers. The specific technical solution adopted is as follows: In a first aspect, the present invention provides a wound image analysis method for assisting in the grading of diabetic foot ulcers, comprising: determining the ulcer lesion region in the input image, and determining the equivalent diameter of the lesion based on the area of ​​the ulcer lesion region; determining a first point set consisting of all tissue sample points to be classified within the ulcer lesion region, and a second point set consisting of all potential convergence locations within the ulcer lesion region; wherein the first point set and the second point set have the same number of points, denoted as the first quantity; constructing a total cost matrix based on the equivalent diameter of the lesion, the first point set, and the second point set; wherein the total cost matrix is ​​used to comprehensively characterize the cost of each tissue sample point to be classified moving to any potential convergence location, and the cost of remaining at the original location; solving the total cost matrix using a preset matching algorithm, and classifying each point in the first point set as a liquid attribute point or a solid attribute point based on the solution result; determining the proportion of liquid components based on the number of points classified as liquid attribute points and the first quantity, and outputting the corresponding diabetic foot ulcer grading result based on the comparison result of the liquid component proportion with a preset threshold.

[0005] Secondly, this invention provides a wound image analysis system for assisting in the grading of diabetic foot ulcers, comprising: a lesion identification module, a point set construction module, a matrix construction module, an attribute segmentation module, and a result output module; the lesion identification module is used to determine the ulcer lesion region in the input image and determine the equivalent diameter of the lesion based on the area of ​​the ulcer lesion region; the point set construction module is used to determine a first point set consisting of all tissue sample points to be classified within the ulcer lesion region, and a second point set consisting of all potential convergence locations within the ulcer lesion region; wherein the first point set and the second point set have the same number of points, denoted as the first number; the matrix construction module, The system is used to construct a total cost matrix based on the equivalent diameter of the lesion, a first set of points, and a second set of points. The total cost matrix comprehensively represents the cost of moving each tissue sample point to be classified to any potential convergence location, as well as the cost of remaining at the original location. The attribute classification module is used to solve the total cost matrix using a preset matching algorithm, and classifies each point in the first set of points into liquid attribute points or solid attribute points based on the solution results. The result output module is used to determine the proportion of liquid components based on the number of points classified as liquid attribute points and a first quantity, and output the corresponding diabetic foot ulcer grading result based on the comparison result of the proportion of liquid components with a preset threshold.

[0006] Thirdly, the present invention provides an electronic device, comprising: a processor and a memory; wherein the memory is used to store one or more programs, the one or more programs including computer-executable instructions, and when the electronic device is running, the processor executes the computer-executable instructions stored in the memory to cause the electronic device to perform the wound image analysis method for assisting in the grading of diabetic foot ulcers as described in the first aspect and any possible implementation thereof.

[0007] This invention offers the following advantages: By simulating the competitive physical processes of pus "flowing" to low-lying areas of the wound under gravity and necrotic tissue "attaching" to its original location due to surface roughness, the visually difficult-to-distinguish problem of classifying pus and necrotic tissue is transformed into a modelable and solvable mathematical problem of optimal transport and matching. First, the wound area is automatically segmented and a normalized scale is determined through color analysis and clustering. Second, two sets of points to be matched are constructed by extracting high-brightness and dark-brightness points. Then, a total cost matrix is ​​constructed to quantify the aforementioned physical competition by calculating the normalized flow distance and texture attachment cost. Next, a global matching algorithm is used to solve the problem, achieving precise attribute classification for each tissue point. Finally, a grading recommendation is output by statistically analyzing the proportion of liquid points and comparing it with clinical thresholds. This method avoids reliance on subjective experience or single visual features. By introducing a physical prior model, it significantly improves the accuracy and robustness of automatically distinguishing between liquid and solid components in diabetic foot ulcer wounds, providing a more reliable objective basis for clinical grading. Attached Figure Description

[0008] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention 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 the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a schematic diagram of the architecture of a wound image analysis system for assisting in the grading of diabetic foot ulcers, provided in one embodiment of the present invention. Figure 2 This is a schematic flowchart of a wound image analysis method for assisting in the grading of diabetic foot ulcers, provided in one embodiment of the present invention. Detailed Implementation

[0010] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the specific implementation methods, structures, features, and effects of the present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0011] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0012] In all division and logarithmic operations involved in this invention, a smoothing mechanism is employed to prevent computer program crashes or invalid values ​​from being generated due to a zero denominator or zero input. Specifically, a correction factor ε, which is a very small positive number, is superimposed on the denominator term of the division operation or the argument term of the logarithmic function, for example, a value of 10 to the power of negative 5, thereby ensuring the robustness and feasibility of the algorithm under extreme conditions.

[0013] Unless otherwise specified, the normalization function Norm() mentioned in this invention uses maximum and minimum value normalization. The maximum and minimum values ​​are preset empirical extreme values ​​derived from a large amount of historical experimental data. If the calculation result exceeds the [0, 1] interval, it is restricted to the [0, 1] range by a truncation function (i.e., if the result is less than 0, it is taken as 0; if it is greater than 1, it is taken as 1) to eliminate the influence of outliers on the evaluation index.

[0014] The following description, in conjunction with the accompanying drawings, details a specific scheme for wound image analysis to assist in the grading of diabetic foot ulcers provided by this invention.

[0015] For example, such as Figure 1The diagram shown is an architectural schematic of a wound image analysis system (hereinafter referred to as the image analysis system) for assisting in the grading of diabetic foot ulcers, according to an embodiment of the present invention. The image analysis system 10 includes: a lesion identification module 11, a point set construction module 12, a matrix construction module 13, an attribute segmentation module 14, and a result output module 15. The modules are described below in sequence: (1) Lesion identification module 11.

[0016] The lesion identification module 11 is responsible for receiving the original wound image and accurately locating the ulcer area through a series of image processing operations. At the same time, it calculates the scale benchmark used for normalization of all subsequent distance calculations, establishing a spatial framework and normalization standard for the entire analysis process.

[0017] Optionally, the lesion identification module 11 is used to determine the ulcer lesion area in the input image and determine the equivalent diameter of the lesion based on the area of ​​the ulcer lesion area.

[0018] Specifically, the lesion recognition module 11 first receives the input color image, which can be acquired by a medical digital camera. Then, the lesion recognition module 11 converts the image from the RGB color space to the CIE LAB color space and performs unsupervised clustering of the image pixels based on the a-channel component values.

[0019] Then, the lesion identification module 11 determines the boundary of the ulcer lesion area based on the regions that meet the preset color characteristics in the clustering results (such as the cluster with the largest a channel value). Finally, the lesion identification module 11 counts the number of pixels in the area to calculate the area, and converts this area into the diameter of an equivalent circle, which is output as the equivalent diameter of the lesion.

[0020] The lesion identification module 11 provides the two key parameters, the boundary information of the ulcer lesion area and the equivalent diameter of the lesion, to the point set construction module 12 simultaneously.

[0021] (2) Point set construction module 12.

[0022] The point set construction module 12 is responsible for intelligently identifying and generating two core point sets within the ulcer lesion area defined by the lesion identification module 11: one type represents visually significant tissues that need to be classified, and the other type represents possible convergence points of fluid. These two types of point sets are the basis for building the competition model in the future.

[0023] Optionally, the point set construction module 12 is used to determine a first point set consisting of all tissue sample points to be classified within the ulcer lesion area, and a second point set consisting of all potential convergence locations within the ulcer lesion area.

[0024] Specifically, the point set construction module 12 first receives the original input image and the boundary of the ulcer lesion region. Then, the point set construction module 12 converts the image to the HSV color space, calculates the specular response map based on brightness and saturation, extracts the bright connected regions through adaptive thresholding, calculates the geometric centroid of each region, and the coordinates of all centroids constitute the first point set.

[0025] Simultaneously, the point set construction module 12 calculates the dark channel map within the ulcer lesion area and determines deep candidate regions based on the color depth threshold. Finally, the point set construction module 12 resamples the deep candidate regions based on the number of points in the first point set, generating a coordinate point set with the exact same number as the first point set, which is then output as the second point set.

[0026] The point set construction module 12 sends the generated first point set and second point set together to the matrix construction module 13.

[0027] (3) Matrix construction module 13.

[0028] The matrix construction module 13 is responsible for receiving all geometric and point set information from the preceding module. By simulating physical processes (flow and attachment), it quantifies the "pairing" cost for each unclassified point and convergence point, as well as each unclassified point itself, thereby constructing a complete and mathematically solvable total cost matrix.

[0029] Optionally, the matrix construction module 13 is used to construct a total cost matrix based on the lesion equivalent diameter, the first point set, and the second point set.

[0030] Specifically, the execution of the matrix construction module 13 includes two parallel cost calculation processes. First, the matrix construction module 13 constructs a surface mesh model based on the ulcer lesion region. For each pair of points in the first and second point sets, the shortest path length is calculated on the mesh, and each path length is normalized by dividing by the equivalent diameter of the lesion. The first cost submatrix is ​​formed by all normalized distance values. This process quantifies the cost of "liquid flow".

[0031] Then, for each point in the first point set, the matrix construction module 13 calculates the gradient variance of its neighborhood image as a local texture roughness index, and determines a cost value representing the tendency of "in-situ attachment" for each point according to the negative correlation mapping principle of "the rougher the texture, the lower the cost value". These cost values ​​form a diagonal second cost submatrix.

[0032] Finally, the matrix construction module 13 concatenates the first cost submatrix and the second cost submatrix to generate a complete total cost matrix containing all possible "matching" options, and outputs it to the attribute partitioning module 14.

[0033] (4) Attribute partitioning module 14.

[0034] The attribute segmentation module 14 is responsible for receiving the total cost matrix generated by the matrix construction module 13, and making a clear attribute decision for each sample point in the first point set by solving a global optimal matching problem, thereby achieving accurate differentiation between liquid and solid states.

[0035] Optionally, the attribute partitioning module 14 is used to solve the total cost matrix using a preset matching algorithm, and classify each point in the first point set into liquid attribute points or solid attribute points according to the solution results.

[0036] Specifically, the attribute partitioning module 14 integrates or calls a preset global optimization matching algorithm, such as the Hungarian algorithm. This algorithm takes the total cost matrix as input and aims to find an optimal set of matching schemes such that each sample point is either uniquely paired with a convergence point or paired with a virtual location representing itself, and the total cost of all pairings is minimized. After running this algorithm, the attribute partitioning module 14 determines the attribute of each sample point based on its final matching object: if it matches a convergence point in a second set of points, the point is determined to be a liquid attribute point; if it matches its own virtual location, the point is determined to be a solid attribute point.

[0037] The attribute segmentation module 14 sends the classification result list to the result output module 15.

[0038] (5) Result output module 15.

[0039] The results output module 15 is responsible for performing statistical calculations based on the classification results of the attribute classification module 14, and mapping the numerical results into ulcer grading suggestions with clinical guidance significance, thus completing the closed loop from image analysis to clinical decision support.

[0040] Optionally, the result output module 15 is used to determine the proportion of liquid components based on the number of liquid attribute points classified as liquid and the first quantity, and to output the corresponding diabetic foot ulcer grading result based on the comparison result of the proportion of liquid components with the preset threshold.

[0041] Specifically, the result output module 15 first receives the classification list sent by the attribute classification module 14. Then, the result output module 15 counts the number of points marked as liquid attribute points in the list. Next, the result output module 15 obtains a first quantity (i.e., the number of points in the first point set) and divides the number of liquid attribute points by this first quantity to calculate the proportion of liquid components.

[0042] Finally, the results output module 15 compares the calculated proportion of liquid components with one or more preset clinical grading thresholds. For example, if the proportion exceeds a high threshold, a grading result of "indicating a risk of deep infection" is generated; if the proportion is below a low threshold, a grading result of "indicating predominantly necrotic tissue adhesion" is generated. This grading result can be displayed through the system's graphical user interface or transmitted to a medical information system in the form of text, charts, or direct annotation and rendering on the original image.

[0043] The above describes the image analysis system 10 and its included modules.

[0044] For example, such as Figure 2 The diagram shown is a flowchart illustrating a wound image analysis method for assisting in the grading of diabetic foot ulcers according to an embodiment of the present invention, comprising the following steps: S201. Determine the ulcer lesion area in the input image, and determine the equivalent diameter of the lesion based on the area of ​​the ulcer lesion region.

[0045] For example, this step can be performed by the lesion identification module 11 in the image analysis system 10 described above, and specifically includes the following steps: (1) Convert the input image to a preset color space and perform clustering processing on the converted image pixels.

[0046] Specifically, the lesion recognition module 11 receives an original RGB image containing a diabetic foot ulcer wound, acquired by an image acquisition device (such as a medical digital camera), as the input image. Then, the lesion recognition module 11 converts the input image from the RGB color space to the CIE LAB color space, which is more suitable for color differentiation and analysis. Finally, based on the a-channel (red-green complementary channel) component values ​​of the converted image in the CIE LAB color space, the lesion recognition module 11 uses a preset unsupervised clustering algorithm (such as K-means clustering) to classify all pixels, grouping pixels with similar color features into the same category.

[0047] (2) Determine the ulcer lesion area based on the area that meets the preset color characteristics in the classification results, and determine the equivalent diameter of the lesion based on the area of ​​the ulcer lesion area.

[0048] Furthermore, the lesion identification module 11 determines the area of ​​the ulcer lesion region based on the number of all pixels constituting the ulcer lesion region, and then uses the diameter of an equivalent circle with the same area as the lesion lesion region as the equivalent diameter of the lesion. Specifically, after completing pixel clustering, the lesion identification module 11 selects the category with the largest a-channel component value at the cluster center, considering that pixels of this category are most likely to correspond to the yellowish-white necrotic tissue or purulent secretions area in the wound. The lesion identification module 11 extracts the largest connected region formed by pixels of this category and determines its boundary coordinate set as the ulcer lesion region. Subsequently, the lesion identification module 11 counts the total number of pixels contained in this region, using this as the area of ​​the ulcer lesion region (in pixels). Finally, the lesion identification module 11 reverse-engineers the formula for the area of ​​a circle to calculate the diameter of a circle with the same area, using this as the normalized scale benchmark for all subsequent distance calculations.

[0049] Therefore, the lesion identification module 11 achieves automatic and accurate segmentation of the ulcer lesion area through color space conversion and cluster analysis, and provides a unified scale normalization standard for subsequent spatial distance calculation by calculating the equivalent diameter of the lesion, thus eliminating the influence of wounds of different sizes on the analysis results.

[0050] S202. Determine a first point set consisting of all tissue sample points to be classified within the ulcer lesion area, and a second point set consisting of all potential convergence locations within the ulcer lesion area. The first point set and the second point set have the same number of points, denoted as the first number.

[0051] For example, this step can be performed by the point set construction module 12 in the image analysis system 10 described above. Specifically, the point set construction module 12 first extracts points representing bright tissue (first point set) and points representing deep depressions (second point set) within the ulcer lesion area determined by the lesion recognition module 11, based on the color and brightness features of the image, ensuring that both point sets contain the same number of points, which is denoted as the first number. It should be noted that the specific process of the aforementioned sub-steps can be found in S301-S305 below, and will not be repeated here. In another possible implementation, when the point set construction module 12 determines the first point set, it can also use a spot detection algorithm based on multi-scale Gaussian difference to detect local brightness maxima within the ulcer lesion area, and use the coordinate set of these maxima as the first point set to capture significantly bright tissue areas at different scales.

[0052] In another possible implementation, when the point set construction module 12 determines the second point set, it can also use a method combining morphological erosion operations with distance transformation. First, the binary mask of the ulcer lesion region is eroded multiple times to simulate the "deep part". Then, the distance transformation map of the eroded region is calculated, and the coordinates of the local maxima points in the distance transformation map are used as potential convergence locations to construct the second point set.

[0053] Thus, the point set construction module 12 successfully abstracts continuous image regions into two discrete point sets of equal quantity by extracting visual high-brightness points to simulate samples to be classified and by sampling dark regions to simulate potential fluid convergence points, laying the foundation for the subsequent construction of a mathematical model based on point pair matching.

[0054] S203. Construct a total cost matrix based on the equivalent diameter of the lesion, the first point set, and the second point set. The total cost matrix comprehensively represents the cost of moving each tissue sample point to any potential convergence location, as well as the cost of remaining at its original location.

[0055] For example, this step can be performed by the matrix construction module 13 in the image analysis system 10 described above. Specifically, the matrix construction module 13 uses the equivalent diameter of the lesion as a normalization factor and the ulcer lesion area as a constraint space to calculate the normalized path cost from each point in the first point set to each point in the second point set, as well as the cost of each point "attaching" to its original location. These two cost components are then combined into a complete square matrix, i.e., the total cost matrix. It should be noted that the specific procedures for the aforementioned sub-steps can be found in S401-S403 below, and will not be repeated here.

[0056] In another possible implementation, when constructing the total cost matrix, the matrix construction module 13 can also directly calculate the two-dimensional Euclidean distance between each pair of points in the first and second point sets, and then normalize it by dividing by the equivalent diameter of the lesion to form the first cost submatrix. This method is simpler to calculate and is suitable for situations where the wound surface is relatively flat.

[0057] Thus, the matrix construction module 13 constructs a mathematical model (total cost matrix) that can simultaneously express the difficulty of moving organization points and the tendency of attachment by quantifying the costs of the two competing physical processes of "flowing" and "attaching", transforming the complex visual discrimination problem into a solvable optimal matching problem.

[0058] S204. Solve the total cost matrix using a preset matching algorithm, and classify each point in the first point set into liquid attribute points or solid attribute points based on the solution results.

[0059] For example, this step can be performed by the attribute segmentation module 14 in the image analysis system 10 described above.

[0060] Specifically, the attribute partitioning module 14 receives the total cost matrix generated by the matrix construction module 13. This matrix is ​​a square matrix of dimension (N+N)×(N+N), where N is the first quantity. The first N rows and the first N columns of the matrix correspond to the cost of flowing from the first set of points (N points) to the second set of points (N points); the first N rows and the last N columns of the matrix form a diagonal matrix, and the diagonal elements correspond to the cost of each point in the first set remaining in place.

[0061] Next, the attribute partitioning module 14 calls a preset global optimal matching algorithm (such as the Hungarian algorithm) to solve the total cost matrix. The goal of this algorithm is to find an optimal "matching" scheme such that each point in the first point set is uniquely assigned to a "target" (which can be a point in the second point set or a virtual attachment point representing itself), and the total cost of all assignments is minimized. After the solution is completed, the attribute partitioning module 14 analyzes the matching results: if a point in the first point set is matched to a point in the second point set, then the point is determined to be a liquid attribute point (simulating pus flowing to a lower position); if it is matched to its own virtual attachment point, then the point is determined to be a solid attribute point (simulating rotting flesh attaching in situ).

[0062] Thus, the attribute segmentation module 14 solves a global optimization problem and makes a clear physical state (liquid / solid) judgment for each highlighted tissue point, thereby achieving automatic and accurate differentiation between pus and necrotic tissue with highly similar visual features.

[0063] S205. Determine the proportion of liquid components based on the number of points classified as liquid attribute points and the first quantity, and output the corresponding diabetic foot ulcer grading result based on the comparison result of the proportion of liquid components with the preset threshold.

[0064] For example, this step can be performed by the result output module 15 in the image analysis system 10 described above.

[0065] Specifically, the result output module 15 receives the classification result list sent by the attribute classification module 14. First, the result output module 15 counts the number of attribute points marked as liquid in the list, denoted as . Next, the result output module 15 obtains the first quantity N (i.e., the total number of points in the first point set) determined in step S202. Then, the result output module 15 calculates the proportion of liquid components. Finally, the results output module 15 compares the calculated proportion of liquid components R with a preset threshold. If R > T, the wound is determined to be dominated by liquid components, indicating a risk of deep abscess or large amounts of purulent discharge, and a high-level risk warning or corresponding Wagner classification recommendation is output (such as indicating that the infection is progressing to deeper layers). If R ≤ T, the wound is determined to be dominated by solid necrotic tissue, and corresponding treatment recommendations are output (such as indicating that debridement and removal of necrotic tissue are required). This classification result can be displayed intuitively through the system's graphical interface or integrated into the electronic medical record report.

[0066] For example, the aforementioned preset threshold can be set based on clinical experience and historical data statistics; for instance, T=0.6. The setting of the preset threshold is mainly based on the pathological differences between "deep abscess" (Wagner grade 3) and "deep ulcer" (Wagner grade 2) in the Wagner classification of diabetic foot ulcers. Its specific value can be determined by collecting a large number of wound image samples with clearly defined clinical classifications, analyzing and calculating the proportion of liquid components in each category using this method, and then determining the optimal diagnostic cutoff value through statistical learning (such as ROC curve analysis) or clinical expert consensus. The aim is to maximize the differentiation between wound states dominated by infected pus and those dominated by necrotic tissue, thereby achieving accuracy and reliability in auxiliary classification diagnosis.

[0067] Thus, the output module 15 transforms the numerical results of the algorithm's judgment into classification suggestions with clear clinical significance, completing the closed loop from image analysis to assisted diagnostic decision-making, and providing doctors with objective and quantitative wound assessment references.

[0068] Based on the above technical solution, this invention transforms the visually difficult-to-distinguish problem of classifying pus and necrotic tissue into a modelable and solvable mathematical problem of optimal transport and matching by simulating the competitive physical processes of pus "flowing" to low-lying areas of the wound under gravity and necrotic tissue "attaching" to its original location due to surface roughness. First, the wound area is automatically segmented and a normalized scale is determined through color analysis and clustering. Second, two sets of points to be matched are constructed by extracting high-brightness and dark-brightness points. Then, a total cost matrix is ​​constructed to quantify the aforementioned physical competition by calculating the normalized flow distance and texture attachment cost. Next, a global matching algorithm is used to solve the problem, achieving accurate attribute classification for each tissue point. Finally, a grading suggestion is output by statistically analyzing the proportion of liquid points and comparing it with clinical thresholds. This method avoids reliance on subjective experience or single visual features. By introducing a physical prior model, it significantly improves the accuracy and robustness of automatically distinguishing between liquid and solid components in diabetic foot ulcer wounds, providing a more reliable objective basis for clinical grading.

[0069] For example, in another wound image analysis method for assisting in the grading of diabetic foot ulcers provided in one embodiment of the present invention, determining a first point set consisting of all tissue sample points to be classified within the ulcer lesion area, and a second point set consisting of all potential convergence locations within the ulcer lesion area, specifically includes the following steps: S301. Determine the specular response map based on the brightness and saturation characteristics of the input image.

[0070] In this step, the point set construction module 12 receives the ulcer lesion region determined by the lesion recognition module 11 and the original input image. First, the point set construction module 12 converts the input image from the RGB color space to the HSV color space. In the HSV space, the V (Value) channel represents brightness, and the S (Saturation) channel represents saturation. Then, the point set construction module 12 calculates the specular response value for each pixel location by weighted fusion of the brightness component V and the saturation component S. One specific implementation uses the following formula to calculate the specular response map:

[0071] in, This represents the specular response value of the pixel at coordinates (x, y) in the image. This represents the lightness component value of the pixel in the HSV color space; This represents the saturation component value of that pixel.

[0072] Understandably, the above formula multiplies the lightness and saturation components pixel by pixel. In wound images, yellowish-white pus or necrotic tissue areas typically have both high lightness (brighter) and high saturation (purer color). Therefore, the multiplication operation amplifies the feature values ​​of such areas, while dark or dull areas have lower response values, thus generating a feature map that highlights the bright, vivid tissue areas.

[0073] S302. Extract the bright connected regions based on the adaptive threshold of the specular response map.

[0074] Furthermore, the point set construction module 12 performs binarization on the calculated specular response map to separate the highlight regions. In this embodiment, Otsu's method is used as the adaptive thresholding algorithm. The point set construction module 12 automatically calculates an optimal global threshold using Otsu's method. .

[0075] Next, the point set construction module 12 binarizes the specular response map according to this threshold: for each pixel position (x, y) in the map, if If a pixel is found to be in the foreground (highlight), it is marked as such; otherwise, it is marked as background. Finally, the point set construction module 12 performs connected component analysis on the binary image, identifying all interconnected foreground pixel sets as independent highlighted connected regions and recording their outer contours.

[0076] S303. Determine the geometric centroid of each highlighted connected region, and define the set of all centroids as the first point set.

[0077] Finally, the point set construction module 12 calculates the geometric centroid coordinates for each highlighted connected region extracted in S302. For a connected region containing N pixels, its centroid coordinates are... Calculated using the following formula:

[0078] in, and These represent the x and y coordinates of the centroid of the connected region, respectively; N represents the total number of pixels contained in the connected region. and They represent the first in this region. i The x and y coordinates of each pixel are calculated. It should be noted that the above formula calculates the arithmetic mean of the coordinates of all pixels within a connected region, using this as the geometric center of that region. The point set construction module 12 calculates the centroid coordinates of all highlighted connected regions and sets these coordinate points together. The first point set is determined, where M is the total number of identified highlighted connected regions.

[0079] S304. Determine the dark channel map within the ulcer lesion area, and determine the pixel areas with color depth lower than the first depth threshold as deep candidate areas based on the dark channel map.

[0080] Specifically, the point set construction module 12 calculates the dark channel map within the ulcer lesion region determined by the lesion identification module 11. First, for each pixel within the ulcer lesion region, the point set construction module 12 takes the minimum value among its three RGB color channels to obtain the dark channel value of that pixel. The calculation formula is as follows:

[0081] in, This represents the dark channel value at coordinates (x, y); Ω represents the intensity value of the pixel in the c color channel (R,G,B); Ω represents a local neighborhood window of the current pixel (e.g., 3x3 or 5x5). In this embodiment, for the sake of simplifying the calculation, Ω may only contain the pixel itself.

[0082] It should be noted that the above formula finds the minimum intensity value of each pixel in the three RGB color channels. In diabetic foot ulcer wounds, areas that are dark in color, appear sunken, or necrotic usually have low values ​​in all RGB channels, and therefore low dark channel values; while bright or vividly colored areas have higher dark channel values. The resulting dark channel map can effectively highlight dark areas in the wound. Subsequently, the point set construction module 12 determines deep candidate regions based on the dark channel map.

[0083] In this embodiment, the first depth threshold The lower quartile (i.e., the 25th percentile) of all pixel values ​​in the dark channel image within the ulcer lesion region can be used as an example. Point set construction module 12 will use the dark channel values... Below this threshold All pixels are extracted, and the connected regions formed by these pixels are identified as deep candidate regions. It should be noted that the first depth threshold selects pixels with color depth in the bottom 25% through statistical distribution, which can reliably capture the darkest parts of the wound, which are more likely to correspond to low-lying areas of fluid accumulation visually and physically.

[0084] S305. Based on the number of points in the first point set and the number of pixels in the deep candidate region, the deep candidate region is resampled to generate a coordinate point set with the same number of points as the first point set to form the second point set.

[0085] Furthermore, the point set construction module 12 performs a resampling operation. First, it obtains the number of points in the first point set, denoted as . Simultaneously, the total number of pixels within the deep candidate region is counted and denoted as... The point set construction module 12 adopts different sampling strategies based on the relationship between these two quantities: like The point set construction module 12 randomly and uniformly extracts pixel coordinates without replacement from all pixel coordinates of the deep candidate region. Coordinates.

[0086] like Point set construction module 12 first sets all the deep candidate regions Select all pixel coordinates. Then, from this... The coordinates are randomly selected with replacement until the total number of selected coordinates reaches a certain threshold. This means that some coordinates may be selected repeatedly.

[0087] Through the above resampling process, the point set construction module 12 finally generates a set containing exactly... The set of coordinate points, regardless of the actual size of the deep candidate region, is consistent with the first set of points in terms of the number of points, and this set is determined as the second set of points.

[0088] Based on the above technical solution, this embodiment of the invention simulates visually appealing samples to be classified by calculating the specular response map and extracting the centroid of the bright tissue region as a first point set using an adaptive threshold. Simultaneously, it simulates potential fluid convergence locations by calculating the dark channel map and locating dark regions using a statistical threshold, and by resampling to generate a second point set that matches the first point set in quantity. This achieves intelligent abstraction from continuous images to discrete, equal-quantity point pairs, providing accurate and reliable input for subsequent construction of a physical competition model based on optimal matching. It is a crucial preprocessing step for distinguishing between liquid and solid components.

[0089] For example, in another wound image analysis method for assisting in the grading of diabetic foot ulcers provided in one embodiment of the present invention, determining a first point set consisting of all tissue sample points to be classified within the ulcer lesion area, and a second point set consisting of all potential convergence locations within the ulcer lesion area, specifically includes the following steps: S401. Determine the cost of moving each sample point of the organization to be classified to each potential convergence location to form a first cost submatrix.

[0090] Optionally, matrix construction module 13 constructs the first cost submatrix through the following sub-steps: (1) Construct a surface mesh model based on the ulcer lesion area.

[0091] Specifically, the matrix construction module 13 constructs a graph structure model, or surface mesh model, to simulate the anatomical surface of the wound, based on the binary mask of the ulcer lesion region determined by the lesion identification module 11. In this model, each pixel within the ulcer lesion region is considered a graph point. The matrix construction module 13 establishes edges between adjacent points based on the 8-neighborhood connectivity of pixels (i.e., top, bottom, left, right, and four diagonal directions). The weight of each edge is set to the physical distance between the centers of two adjacent pixels (e.g., 1 pixel unit for horizontal or vertical adjacency, and √2 pixel units for diagonal adjacency). Pixels that do not belong to the ulcer lesion region are not included in this mesh model and are considered impassable obstacles during path calculation. This model simulates the physical constraint that liquids can only flow on the continuous surface of wound tissue and cannot penetrate healthy tissue.

[0092] (2) For each point in the first point set and each point in the second point set, calculate the shortest path length between the two points on the surface mesh model.

[0093] Specifically, matrix construction module 13 iterates through each point in the first point set (denoted as P, containing N points). With each point in the second point set (denoted as Q, containing N points) For each pair of points Matrix construction module 13 runs a shortest path search algorithm on the surface mesh model constructed in step (1) to calculate the path from point to point. Time The path. In this embodiment, Dijkstra's algorithm is used as the shortest path search algorithm. If point Time For a surface mesh model that is connected (i.e., there exist vertex paths connected by edges), Dijkstra's algorithm returns the cumulative length of the shortest path between them, denoted as . (In pixels). If two points are disconnected due to a break in the model (such as being isolated by a healthy skin island), the matrix construction module 13 will calculate the path length. The value is assigned to a preset maximum distance value. ,in It can be set to 10 times the equivalent diameter of the lesion ( ),Right now This is used mathematically to represent "flow is unreachable", which would have extremely high costs.

[0094] (3) Construct the first cost submatrix based on the shortest path length and the equivalent diameter of the lesion.

[0095] Finally, the matrix construction module 13 utilizes the lesion equivalent diameter provided by the lesion identification module 11. The original shortest path length is calculated for each pair of points. Normalization was performed to eliminate the influence of different wound sizes on the distance metric. The normalized distance values ​​constituted the first cost submatrix. The element in the i-th row and j-th column. The specific calculation formula is:

[0096] in, This represents the normalized flow cost corresponding to the i-th point of the first point set and the j-th point of the second point set in the first cost submatrix; This represents the points calculated on the surface mesh model. Time The shortest path length (in pixels); Indicates the equivalent diameter of the lesion (in pixels).

[0097] It should be noted that the above formula divides the original path length by the characteristic scale of the wound (the equivalent diameter of the lesion), yielding a dimensionless relative distance. This value characterizes the "path resistance" that fluid must overcome to "flow" from a highlighted tissue sample point to a deep potential convergence point at the "current wound scale." All N×N such normalized distance values ​​form an N x N square matrix, which is the first cost submatrix. .

[0098] S402. Determine the cost of each tissue sample point remaining in its original position to form a second cost submatrix.

[0099] Optionally, matrix construction module 13 constructs the second cost submatrix through the following sub-steps: (1) For each point in the first point set, the local texture roughness index is determined based on the gradient features of the neighborhood image.

[0100] Specifically, the matrix construction module 13 targets each point in the first point set P. The following operations are performed to quantify the coarseness of its local texture. First, the matrix construction module 13 converts the original input image into a grayscale image and computes its gradient magnitude map (e.g., using the Sobel operator). Then, using points... Centered on the coordinates of the point, a fixed-size local neighborhood window, such as a 5-pixel × 5-pixel square, is extracted from the gradient magnitude map. Then, the variance of the gradient magnitudes of all pixels within this neighborhood window is calculated and denoted as the initial roughness metric for that point. To perform a relative comparison, matrix construction module 13 iterates through all N points in the first point set to find the largest initial roughness value, denoted as . Finally, the local texture roughness index for each point is calculated using the following formula. :

[0101] in, This represents the local texture roughness index of the i-th point in the first point set, and its value is between 0 and 1. This represents the variance of the gradient magnitude within the neighborhood centered at point p_i; This represents the largest gradient magnitude variance among all points.

[0102] Understandably, the above formula compares and normalizes the local gradient variance at each point with the global maximum variance. A larger gradient variance indicates more drastic texture variations around that point, corresponding to a rougher microscopic surface of the tissue (e.g., the fibrous structure of solid putrefied flesh); a smaller gradient variance indicates a smoother texture (e.g., the surface of liquid pus). Therefore, The larger the value, the coarser the texture of the tissue sample corresponding to that point.

[0103] (2) Construct a second submatrix based on the local texture roughness index of each point.

[0104] Furthermore, the matrix construction module 13 inversely maps the local texture roughness index ri to the cost of "remaining in place." This is to simulate the physical law: the rougher the texture (the stronger the solid-state properties), the stronger the tendency of the tissue to adhere to the wound substrate; therefore, the cost of the decision to "choose not to flow and remain in place" should be lower. In this embodiment, the adhesion cost for each point is calculated using the following formula. :

[0105] in, This represents the in-situ attachment cost of the i-th point in the first point set; It is a preset balance coefficient used to adjust the order of magnitude of attachment cost and flow cost in the total cost. In this embodiment, it can be set to 0.5.

[0106] It should be noted that the attached cost in the above formula Roughness index They show a negative correlation. When When it approaches 1 (extremely coarse), Approaching 0, making Very small, indicating low cost of "attachment"; when When it approaches 0 (extremely smooth), Approaching 1, Approximately α, indicating a high cost for "attachment". The attachment cost at all N points. This forms a vector of length N. Matrix construction module 13 uses this vector to construct an N x N diagonal matrix as the second submatrix. That is, the elements on the main diagonal of the matrix are, in order. Elements off-diagonal are given a very large penalty value (e.g., 2×). , (This is a preset maximum distance value) to ensure that in subsequent matching solutions, each point can only match its own corresponding attachment option, and cannot be incorrectly matched to the attachment options of other points.

[0107] S403. Generate the total cost matrix based on the first cost submatrix and the second cost submatrix.

[0108] For example, matrix construction module 13 will convert the first cost submatrix generated in S401 into a matrix. The second cost submatrix generated by S402 Horizontal stitching is performed to generate the final total cost matrix. . Specifically, It is a matrix with dimensions N rows × 2N columns. Its left half (columns 1 to N) is placed directly... Its right half (columns N+1 to 2N) is placed The mathematical form is as follows:

[0109] The above formula represents the total cost matrix. From the first cost quantum moment Second submatrix They are arranged side by side. The significance of this construction is that it represents each point in the first set of points (N sample points to be classified). It defines 2N possible "destination" choices and their corresponding costs: the first N choices are to flow to N potential convergence points in the second point set. The cost is The last N choices are to stay in place (i.e., attach), but in order to constitute a perfect matching problem, each point... It can only match its exclusive (N+i)th attachment option, with a cost of (Right now (Diagonal elements), while the cost of the attachment option to match other points is infinite (by... (This is reflected in the off-diagonal maxima). This total cost matrix... The input to the subsequent global optimal matching problem is fully defined.

[0110] Based on the above technical solution, this embodiment of the invention constructs a bipartite graph matching model based on physical rheology priors by separately calculating the normalized geodesic distance cost simulating the "flowing" path of liquid and the reverse texture roughness cost simulating the "attachment" characteristics of solid, and integrating the two into an augmented total cost matrix. This model transforms the visual challenge of distinguishing pus from necrotic tissue into a mathematical optimization problem of selecting the path with the lowest global cost for each sample point in both flowing and attaching behaviors, laying a core mathematical model foundation for subsequent accurate classification using matching algorithms.

[0111] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0112] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A wound image analysis method for assisting in the grading of diabetic foot ulcers, characterized in that, The method includes: Identify the ulcer lesion area in the input image, and determine the equivalent diameter of the lesion based on the area of ​​the ulcer lesion region; A first set of points consisting of all tissue sample points to be classified within the ulcer lesion area is determined, and a second set of points consisting of all potential convergence locations within the ulcer lesion area is determined; wherein, the first set of points and the second set of points have the same number of points, denoted as the first number; Based on the equivalent diameter of the lesion, the first point set, and the second point set, a total cost matrix is ​​constructed; wherein, the total cost matrix is ​​used to comprehensively characterize the cost of each of the tissue sample points to be classified moving to any of the potential convergence locations, as well as the cost of remaining at the original location; The total cost matrix is ​​solved using a preset matching algorithm, and each point in the first point set is classified as a liquid attribute point or a solid attribute point based on the solution result. The proportion of liquid components is determined based on the number of points classified as liquid attribute points and a first quantity, and the corresponding diabetic foot ulcer grading result is output based on the comparison result of the liquid component proportion with a preset threshold.

2. The wound image analysis method for assisting in the grading of diabetic foot ulcers according to claim 1, characterized in that, Identify the ulcer lesion region in the input image, and determine the equivalent diameter of the lesion based on the area of ​​the ulcer lesion region, specifically including: The input image is converted to a preset color space, and the pixels of the converted image are clustered. The ulcer lesion area is determined based on the area that matches the preset color characteristics in the classification results, and the equivalent diameter of the lesion is determined based on the area of ​​the ulcer lesion area.

3. The wound image analysis method for assisting in the grading of diabetic foot ulcers according to claim 2, characterized in that, The equivalent diameter of the lesion is determined based on the area of ​​the ulcer lesion region, specifically including: The area of ​​the ulcer lesion region is determined based on the number of all pixels constituting the ulcer lesion region; The diameter of the equivalent circle with the same area as the ulcer lesion is taken as the equivalent diameter of the lesion.

4. The wound image analysis method for assisting in the grading of diabetic foot ulcers according to claim 1, characterized in that, The first point set, comprising all unclassified tissue sample points within the ulcer lesion area, is determined, specifically including: Based on the brightness and saturation characteristics of the input image, a specular response map is determined; Based on the adaptive threshold of the specular response map, extract the bright connected regions; Determine the geometric centroid of each of the highlighted connected regions, and define the set of all centroids as the first point set.

5. The wound image analysis method for assisting in the grading of diabetic foot ulcers according to claim 1, characterized in that, Determine the second set of points comprising all potential convergence locations within the ulcer lesion area, specifically including: A dark channel map is determined within the ulcer lesion area, and pixel regions with color depths lower than a first depth threshold are identified as deep candidate regions based on the dark channel map. Based on the number of points in the first point set and the number of pixels in the deep candidate region, the deep candidate region is resampled to generate a coordinate point set with the same number of points as the first point set to form the second point set.

6. The wound image analysis method for assisting in the grading of diabetic foot ulcers according to claim 1, characterized in that, Based on the equivalent diameter of the lesion, the first point set, and the second point set, a total cost matrix is ​​constructed, specifically including: Determine the cost of moving each of the tissue sample points to be classified to each of the potential convergence locations to form a first cost submatrix; Determine the cost of each of the tissue sample points to be classified remaining at its original position to form a second cost submatrix; The total cost matrix is ​​generated based on the first cost submatrix and the second cost submatrix.

7. The wound image analysis method for assisting in the grading of diabetic foot ulcers according to claim 6, characterized in that, Determine the cost of moving each of the tissue sample points to be classified to each of the potential convergence locations to form a first cost submatrix, specifically including: A surface mesh model is constructed based on the ulcer lesion area; For each point in the first set of points and each point in the second set of points, calculate the shortest path length between the two points on the surface mesh model; The first cost submatrix is ​​constructed based on each of the shortest path lengths and the equivalent diameter of the lesion.

8. The wound image analysis method for assisting in the grading of diabetic foot ulcers according to claim 6, characterized in that, Determine the cost of each of the tissue sample points to be classified remaining at its original position to form a second cost submatrix, specifically including: For each point in the first point set, a local texture roughness index is determined based on the gradient features of the neighborhood image. The second cost submatrix is ​​constructed based on the local texture roughness index of each point.

9. The wound image analysis method for assisting in the grading of diabetic foot ulcers according to any one of claims 1-8, characterized in that, The types of preset matching algorithms include global optimization algorithms.

10. The wound image analysis method for assisting in the grading of diabetic foot ulcers according to any one of claims 1-8, characterized in that, Before determining the ulcer lesion area in the input image, the method further includes: acquiring a raw color image containing the diabetic foot ulcer wound as the input image.

Images