A pressure damage detection method based on fractional-order composite multi-scale sample entropy

Abstract

The present invention belongs to the field of medical image analysis and relates to a pressure injury detection method based on fractional-order composite multi-scale sample entropy, comprising: dynamically tracking a region of interest in a pressure injury video based on an adaptive optical flow field to obtain a tracking video of the region of interest; processing the tracking video using an Euler video magnification method to construct a brightness image sequence; performing fractional-order differential processing on the brightness image sequence to calculate multi-scale sample entropy eigenvalues, and constructing a sample entropy eigenvector based on the multi-scale sample entropy eigenvalues; constructing a training sample set based on the sample entropy eigenvector to train a support vector machine classifier to obtain a pressure injury staging model; and inputting a pressure injury video to be tested into the pressure injury staging model to obtain an injury staging result for the pressure injury video to be tested, thereby improving the accuracy and efficiency of pressure injury detection.

Description

Technical Field

[0001] The present invention relates to the field of medical image analysis, and specifically discloses a pressure injury detection method based on fractional-order composite multi-scale sample entropy. Background Art

[0002] Pressure injuries, also known as pressure ulcers, occur due to prolonged localized pressure on tissue, leading to impaired blood circulation and, subsequently, ischemia, hypoxia, and malnutrition in the skin and subcutaneous tissue. They are common in patients who are bedridden or wheelchair-bound, and are more likely to occur over bony prominences such as the sacrum and heels. Early identification and staging of pressure injuries are crucial to preventing their progression, but existing detection methods still have many limitations.

[0003] Current clinical testing primarily relies on static image recognition techniques, such as erythema threshold segmentation based on the HSV color model, which can identify erythema areas and, combined with morphological processing, locate suspected pressure injuries. Alternatively, gray-level co-occurrence matrices (GLCMs) can be used to quantify the roughness and uniformity of skin texture to distinguish normal tissue from damaged areas. Furthermore, near-infrared spectroscopy or thermal imaging techniques are also used to determine local ischemia based on temperature differences. However, these methods are based on single-frame image analysis and can only capture transient static features, failing to reflect the dynamic evolution of pressure injuries. For example, the early erythema of stage I pressure injuries may exhibit periodic brightness changes with microcirculatory fluctuations, making static images prone to misinterpretation as temporary congestion. Traditional image entropy methods are sensitive to lighting conditions and exhibit significantly reduced specificity in patients with significant skin color variations. Although existing research has attempted to incorporate time-series analysis methods such as multiscale entropy, these methods employ fixed parameters and a coarse-grained mean value. This approach fails to effectively preserve high-frequency signal details (such as transient fluctuations in erythema and is insensitive to subtle dynamic changes), and also struggles to adaptively adjust parameters to accommodate individual skin characteristics. This limits the accuracy of pressure injury staging. Therefore, a method integrating dynamic signal analysis is urgently needed to address the shortcomings of static image technology and achieve accurate staging and real-time monitoring of pressure injuries.

[0004] In view of this, the present invention provides a pressure injury detection method based on fractional composite multiscale entropy (FrCMSE) to improve the accuracy and efficiency of pressure injury detection. Summary of the Invention

[0005] The object of the present invention is to provide a pressure injury detection method based on fractional-order composite multi-scale sample entropy. The specific scheme includes: dynamically tracking the area of ​​interest of the pressure injury video based on the adaptive optical flow field to obtain the tracking video of the area of ​​interest; processing the tracking video by the Euler video magnification method to construct a brightness image sequence; performing fractional-order differential processing on the brightness image sequence to calculate the multi-scale sample entropy eigenvalues, and constructing a sample entropy feature vector based on the multi-scale sample entropy eigenvalues; constructing a training sample set based on the sample entropy feature vector to train a support vector machine classifier to obtain a pressure injury staging model; inputting the pressure injury video to be tested into the pressure injury staging model to obtain the injury staging result of the pressure injury video to be tested.

[0006] Furthermore, the dynamic tracking of the region of interest in the pressure injury video based on the adaptive optical flow field includes: determining the image brightness variance based on the degree of illumination change of the video frame image; determining the image gradient mean based on the motion complexity of the video frame image; determining the adaptive weight coefficient of the adaptive optical flow field based on the image brightness variance and the image gradient mean; and constructing an optimization objective function of the brightness consistency error of the adaptive optical flow field based on the adaptive weight coefficient.

[0007] Furthermore, the calculation formula of the adaptive weight coefficient is:

[0008] ;

[0009] The calculation formula of the optimization objective function of the brightness consistency error is:

[0010] ;

[0011] in, represents the adaptive weight coefficient; and Represent the brightness weight coefficient and gradient weight coefficient respectively; represents the image brightness variance of frame images in the pressure injury video; Represents the image gradient mean of the frame image in the pressure injury video; Indicates the brightness value of the pressure injury video; represents the spatial gradient vector of brightness; represents the gradient modulus; represents the value of the optimization objective function of brightness consistency error; i represents the pixel variable in the region of interest; N represents the total number of pixels in the region of interest; Represents the brightness gradient of the frame image in the pressure injury video in the x direction; Represents the brightness gradient of the frame image in the pressure injury video in the y direction; Represents the temporal gradient of the brightness of the frame image in the pressure injury video; Represents the displacement of the i-th pixel variable in the x direction; Represents the displacement of the i-th pixel variable in the y direction; Represents the displacement gradient of the i-th pixel variable in the x direction; Represents the displacement gradient of the i-th pixel variable in the y direction.

[0012] Furthermore, the fractional-order differential processing is performed on the brightness image sequence, multi-scale sample entropy eigenvalues ​​are calculated, and a sample entropy eigenvector is constructed based on the multi-scale sample entropy eigenvalues, including: nonlinearly enhancing the brightness image sequence through fractional-order differentials to obtain fractional-order derivatives at multiple time points, arranging the fractional-order derivatives to obtain a fractional-order derivative sequence; coarsening the fractional-order derivative sequence based on multiple scale factors to obtain a multi-scale coarse-grained sequence; and performing multi-scale sample entropy calculation on the multi-scale coarse-grained sequence to obtain the sample entropy eigenvector.

[0013] Furthermore, the nonlinear enhancement of the brightness image sequence by fractional-order differentiation is performed to obtain the calculation formula of the fractional-order derivatives at multiple time points:

[0014] ;

[0015] in, Indicates at a point in time The fractional derivative at ; t represents the time variable; Represents the brightness image sequence at time point t; It indicates the value of the calculation formula when the sampling time interval approaches 0 from a direction greater than 0; h indicates the sampling time interval; represents the fractional order; k represents the length of the backtracking time; Represents the lower limit parameter; represents the generalized binomial coefficient; Indicates a time point Brightness image sequence at .

[0016] Furthermore, the coarse-graining of the fractional-order derivative sequence based on multiple scale factors to obtain a coarse-grained sequence includes: determining multiple scale factors, and dividing the fractional-order derivative sequence into multiple non-overlapping time windows of multiple time scales according to the time lengths of the multiple scale factors; calculating the average value of the fractional-order derivative in each non-overlapping time window to obtain coarse-grained data points; arranging the coarse-grained data points of each scale factor to obtain a coarse-grained sequence; and arranging the coarse-grained sequences of multiple scales to obtain a multi-scale coarse-grained sequence.

[0017] Furthermore, the multi-scale sample entropy calculation is performed on the multi-scale coarse-grained sequence to obtain the sample entropy feature vector, including: calculating the sample entropy of each coarse-grained sequence in the multi-scale coarse-grained sequence respectively, including: constructing an embedding vector of a first dimension to obtain a first number of first templates; calculating the distance between the first templates, and taking the number of first templates whose distance is less than the distance tolerance as a first numerical parameter; constructing an embedding vector of a second dimension to obtain a second number of second templates; calculating the distance between the second templates, and taking the number of second templates whose distance is less than the distance tolerance as a second numerical parameter; calculating the sample entropy based on the first numerical parameter and the second numerical parameter; calculating the composite sample entropy under multiple scale factors based on the sample entropy; and arranging the composite sample entropy to obtain the sample entropy feature vector.

[0018] Furthermore, it also includes weighted processing of the composite sample entropy under multiple scale factors in the sample entropy feature vector to obtain a weighted sample entropy feature vector, including: dividing the value range of the composite sample entropy to obtain multiple pressure injury categories; based on the value of the composite sample entropy, respectively determining the target pressure injury category into which the multiple composite sample entropies fall; based on the pressure injury category and the target pressure injury category, calculating the correlation between the composite sample entropy and the multiple pressure injury categories; based on the target pressure injury category, respectively determining the redundancy between the composite sample entropies; based on the correlation and its corresponding redundancy, calculating a weight score; normalizing the weight score to obtain a weighted weight of each composite sample entropy; taking the product of the weighted weight and the corresponding composite sample entropy as the weighted sample entropy, arranging the weighted sample entropies to obtain a weighted sample entropy feature vector.

[0019] Furthermore, the calculation formula of the correlation is:

[0020] ;

[0021] in, represents the first composite sample entropy Correlation with multiple pressure injury categories separately; Indicates the target pressure injury category Belongs to the target pressure injury category to which the first composite sample entropy belongs; Indicates the Scaled composite sample entropy; represents the first composite sample entropy variable; Indicates that pressure injury category c belongs to multiple pressure injury categories ; Represents a set of pressure injury categories, including multiple pressure injury categories; represents the joint probability of the target pressure injury category a and the pressure injury category c of the first composite sample entropy; represents the first composite sample entropy The target pressure injury category is The marginal probability of represents the marginal probability of pressure injury category c;

[0022] The calculation formula of the redundancy is:

[0023] ;

[0024] in, represents the first composite sample entropy Redundancy; represents the second composite sample entropy variable; Indicates that the target pressure injury category b belongs to the target pressure injury category to which the second composite sample entropy belongs; represents the first composite sample entropy The target pressure injury category is And the second composite sample entropy The joint probability that the target pressure injury category is b; represents the second composite sample entropy The marginal probability that the target pressure injury category is b;

[0025] The calculation formula of the weight score is:

[0026] .

[0027] Furthermore, the training sample set based on the sample entropy feature vector is constructed to train a support vector machine classifier to obtain a pressure injury staging model, including:

[0028] Constructing the training sample set based on the sample entropy feature vector;

[0029] The radial basis function is used as the kernel function of the support vector machine classifier; the expression of the kernel function is:

[0030] ;

[0031] in, represents the kernel function; Indicates the weighted sample entropy feature vector; Indicates the weighted sample entropy feature vector; exp represents the exponential function; represents the kernel function parameter; ‖*‖ represents the Euclidean norm;

[0032] When training the support vector machine classifier, the optimal hyperplane is found by optimizing the model objective function to obtain the pressure injury staging model; the expression of the model objective function is:

[0033] ;

[0034] Where, Indicates finding the minimum value of the hyperplane normal vector W and the bias term d; W represents the hyperplane normal vector; d represents the bias term; represents the regularization parameter; q represents the training sample variable; Q represents the total number of training samples; represents the slack variable.

[0035] The present invention preprocesses the input optical flow tracking ROI video data through Euler video amplification to enhance the patient's skin micro-vibration signal and capture the slight changes in the brightness signal.

[0036] The present invention performs fractional-order differential processing on the amplified ROI image sequence, calculates its FrCMSE value, and extracts the FrCMSE eigenvector that reflects the dynamic characteristics of pressure injuries. Fractional-order differentials are an extension of traditional integer-order calculus, allowing the differential order to be any real number, and can better describe signals with nonlinear and nonstationary characteristics. In pressure injury video detection, fractional-order differentials perform nonlinear enhancement on the ROI (region of interest) image sequence, highlighting high-frequency details (such as skin texture and subtle changes in early pressure injuries) while nonlinearly retaining low-frequency information (such as grayscale gradients in smooth skin areas). This processing effectively captures the dynamic characteristics of pressure injuries while suppressing noise interference.

[0037] The present invention calculates the weight of each FrCMSE feature using a minimal redundancy (mRMR) algorithm, maximizing the correlation between the feature and the target category and improving recognition accuracy. To facilitate the calculation of the mutual information between the feature and the category label, the present invention bins (discretizes) the extracted FrCMSE feature vector sample set. Considering that FrCMSE features are continuous-valued variables, directly participating in the mutual information calculation will lead to unstable probability estimation, especially when the sample size is limited. Therefore, a discretization method based on interval partitioning is adopted to map the continuous feature values ​​into a finite number of discrete states, thereby facilitating the estimation of the probability distribution and subsequent statistical operations.

[0038] This paper uses the processed WFrCMSE feature vector sample set to train an SVM (support vector machine) classifier to construct a pressure injury staging model. SVM achieves classification by finding an optimal hyperplane, maximizing the margin between classes and thus improving the classifier's generalization ability. For pressure injury detection, the radial basis function (RBF) kernel is selected as the SVM kernel. The RBF kernel is effective for nonlinear classification problems and is particularly suitable for data classification in high-dimensional feature spaces. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is an exemplary flow chart of a pressure injury detection method based on fractional-order composite multi-scale sample entropy proposed in the present invention. DETAILED DESCRIPTION

[0040] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0041] Figure 1 This is an exemplary flow chart of a pressure injury detection method based on fractional-order composite multi-scale sample entropy proposed by the present invention. Figure 1 As shown in Figure 2, the pressure injury detection method based on fractional-order composite multi-scale sample entropy includes the following:

[0042] Based on the adaptive optical flow field, the region of interest of the pressure injury video is dynamically tracked to obtain a tracking video of the region of interest. In practical applications, the brightness consistency constraint may fail due to factors such as lighting changes and noise interference. In order to improve the robustness of the optical flow field calculation, the parameters of the optical flow calculation are dynamically adjusted according to the motion complexity and lighting conditions of the video frame through the adaptive optical flow field optimization algorithm. The region of interest can refer to the area in the video where pressure damage may exist. The tracking video refers to the video obtained by tracking the region of interest.

[0043] In some embodiments, the adaptive optical flow field optimization algorithm dynamically adjusts the weight coefficient of the optical flow field calculation by analyzing the image brightness variance and gradient information to better adapt to the needs of motion tracking, including: determining the image brightness variance based on the degree of illumination change of the video frame image. Determining the image gradient mean based on the motion complexity of the video frame image. The video frame image refers to the frame image in the pressure injury video. Based on the image brightness variance and the image gradient mean, the adaptive weight coefficient of the adaptive optical flow field is determined. In some embodiments, the calculation formula of the adaptive weight coefficient is:

[0044] ;

[0045] in, Represents the adaptive weight coefficient, which is used to adjust the balance between the smoothing term and the data term in the optical flow field calculation; and Represents the brightness weight coefficient and gradient weight coefficient respectively, which are used to balance the influence of brightness variance and gradient mean on the calculation of optical flow field; The image brightness variance of the frame images in the pressure injury video reflects the degree of change in image brightness. The larger the brightness variance, the more obvious the illumination change in the image, and the stronger the smoothness required for the optical flow field calculation; The mean gradient of the image frame in the pressure injury video reflects the strength of the edges and textures in the image. The larger the mean gradient, the more complex the motion in the image, and the higher the precision required for optical flow calculation. Indicates the brightness value of the pressure injury video; represents the spatial gradient vector of brightness; represents the gradient modulus.

[0046] Based on the adaptive weight coefficient, an optimization objective function of the brightness consistency error of the adaptive optical flow field is constructed. The calculation formula of the optimization objective function of the brightness consistency error is:

[0047] ;

[0048] in, represents the value of the optimization objective function of brightness consistency error; i represents the pixel variable in the region of interest; N represents the total number of pixels in the region of interest; Represents the brightness gradient of the frame image in the pressure injury video in the x direction; Represents the brightness gradient of the frame image in the pressure injury video in the y direction; Represents the temporal gradient of the brightness of the frame image in the pressure injury video; Represents the displacement of the i-th pixel variable in the x direction; Represents the displacement of the i-th pixel variable in the y direction; Represents the displacement gradient of the i-th pixel variable in the x direction; Represents the displacement gradient of the i-th pixel variable in the y direction. and is the smoothing term

[0049] The tracking video is processed by the Euler video magnification method to construct a brightness image sequence. The brightness image sequence refers to an image sequence related to the brightness of the image obtained after processing by the Euler method.

[0050] Performing fractional differential processing on the brightness image sequence, calculating multi-scale sample entropy eigenvalues, and constructing a sample entropy eigenvector based on the multi-scale sample entropy eigenvalues. In some embodiments, performing fractional differential processing on the brightness image sequence, calculating multi-scale sample entropy eigenvalues, and constructing a sample entropy eigenvector based on the multi-scale sample entropy eigenvalues ​​includes: performing nonlinear enhancement on the brightness image sequence by fractional differentials to obtain fractional derivatives at multiple time points, and arranging the fractional derivatives to obtain a sequence of fractional derivatives. In some embodiments, the calculation formula for performing nonlinear enhancement on the brightness image sequence by fractional differentials to obtain fractional derivatives at multiple time points is:

[0051] ;

[0052] in, Indicates at a point in time The fractional derivative at ; t represents the time variable; Represents the brightness image sequence at time point t; Indicates the value of the calculation formula when the sampling time interval approaches 0 from a direction greater than 0; h represents the sampling time interval, which is taken in this method. ; Indicates the fractional order, which controls the signal strength; k indicates the length of the backtracking time, and its value range is In this method, k=30, where K represents the total length of the signal sequence; Represents the lower limit parameter; Indicates a time point Brightness image sequence at ; represents the generalized binomial coefficient, which is calculated by the recursive formula:

[0053] .

[0054] The fractional-order derivative sequence is coarse-grained based on multiple scale factors to obtain a multi-scale coarse-grained sequence. ) Definition: Indicates the granularity of analysis. For example, when =3, the original time series is divided into multiple non-overlapping windows of length 3. Coarse-grained sequence generation: calculate the average value of the data in each window to generate a new coarse-grained sequence. For example, the original sequence 2,4,6,8,10,122,4,6,8,10,12 =2, it becomes 3,7,113,7,11, that is, the average of every two data points is taken.

[0055] In some embodiments, the coarse-graining of the fractional derivative sequence based on multiple scale factors to obtain a coarse-grained sequence includes: determining multiple scale factors, and dividing the fractional derivative sequence into multiple non-overlapping time windows of multiple time scales according to the time lengths of the multiple scale factors. For example, the fractional derivative sequence can be decomposed into multiple time windows at fixed time intervals. The average value of the fractional derivative in each non-overlapping time window is calculated to obtain a coarse-grained data point. The coarse-grained data points of each scale factor are arranged to obtain a coarse-grained sequence. For example, for a time series , in the scale factor The coarse-grained data points under Defined as:

[0056] ;

[0057] Where j is the sequence number after coarse-graining, i represents the order of the original time series before coarse-graining; T is the length of the original sequence; Represents the i-th data point of the original time series; is the scale factor, the final coarse-grained sequence It can be expressed as .

[0058] Arranging coarse-grained sequences of multiple scales to obtain a multi-scale coarse-grained sequence. Performing multi-scale sample entropy calculation on the multi-scale coarse-grained sequence to obtain the sample entropy feature vector includes: calculating the sample entropy of each coarse-grained sequence in the multi-scale coarse-grained sequence, including: constructing an embedding vector of a first dimension to obtain a first number of first templates. For example, constructing an embedding vector of dimension m:

[0059] ;

[0060] There are Nm templates in total.

[0061] Calculate the distance between the first templates and count the number of first templates whose distance is less than the distance tolerance as the first parameter. For example, calculate the distance between templates and count the number of templates whose distance is less than the tolerance r (usually r = 0.15 × standard deviation) as B.

[0062] ;

[0063] statistics The number of is denoted as B.

[0064] Construct an embedding vector of a second dimension to obtain a second number of second templates. Calculate the distance between the second templates, and use the number of second templates whose distance is less than the distance tolerance as a second numerical parameter. For example, if the embedding vector dimension becomes m+1, repeat the above steps to determine the number of templates, and record the number as A.

[0065] The sample entropy is calculated based on the first numerical parameter and the second numerical parameter; the calculation formula for calculating the sample entropy based on the first numerical parameter and the second numerical parameter is:

[0066] ;

[0067] in, represents the sample entropy of the m-dimensional embedding vector with a distance tolerance of r; A represents the second numerical parameter; B represents the first numerical parameter. If the embedding vector dimension increases and remains similar , indicating that the sequence is stable and the entropy is low. If the embedding vector dimension increases, the number of matches will drop significantly. , indicating that the sequence is very unstable, with high entropy and complexity.

[0068] Based on the sample entropy, the composite sample entropy under multiple scale factors is calculated respectively; the calculation formula of the composite sample entropy is:

[0069] ;

[0070] in, Indicates the Composite sample entropy of the scale factor; represents the scale factor variable; Indicates the maximum value of the scale factor variable; Represents a coarse-grained sequence The sample entropy of The scale factor is The coarse-grained sequence of .

[0071] The compound sample entropy is arranged to obtain the sample entropy feature vector. The calculation process of FrCMSE is as follows: first, the brightness signal is subjected to fractional differential processing, and then coarse-grained processing is performed. Calculate the compound sample entropy under , the continuous The sample entropy calculated at each scale finally constructs the FrCMSE feature vector. The calculation formula of the sample entropy feature vector is:

[0072] ;

[0073] in, represents the sample entropy feature vector; 、 and Represent the first scale factor, the second scale factor and the Composite sample entropy of the scale factor.

[0074] In some embodiments, the method further includes performing weighted processing on the composite sample entropy under multiple scale factors in the sample entropy feature vector to obtain a weighted sample entropy feature vector, including:

[0075] The value range of the composite sample entropy is divided to obtain multiple pressure injury categories; based on the value of the composite sample entropy, the target pressure injury category into which the multiple composite sample entropies fall is determined. Feature Dimension First, count the minimum and maximum values ​​in the entire sample set and calculate the feature value range. Then divide the range evenly into V sub-intervals (usually V = 5~10), each sub-interval corresponds to a discrete value number For any sample’s eigenvalue of a certain dimension, determine which interval it belongs to and map it to a corresponding discrete number.

[0076] Based on the pressure injury category and the target pressure injury category, the correlation between the composite sample entropy and multiple pressure injury categories is calculated; the correlation between the pressure injury classification label and the single-dimensional data of the feature vector is calculated, and the calculation formula of the correlation is:

[0077] ;

[0078] in, represents the first composite sample entropy Correlation with multiple pressure injury categories separately; Indicates the target pressure injury category Belongs to the target pressure injury category to which the first composite sample entropy belongs; Indicates the Scaled composite sample entropy; represents the first composite sample entropy variable; Indicates that pressure injury category c belongs to multiple pressure injury categories ,There are 5 categories of pressure injury detection in this task (healthy, pressure injury stage 1, pressure injury stage 2, pressure injury stage 3, pressure injury stage 4); Represents a set of pressure injury categories, including multiple pressure injury categories; represents the joint probability of the target pressure injury category a and the pressure injury category c of the first composite sample entropy; represents the first composite sample entropy The target pressure injury category is The marginal probability of represents the marginal probability of pressure injury category c.

[0079] Based on the target pressure injury category, the redundancy between the composite sample entropies is determined respectively; duplicate information between features is eliminated to ensure strong independence of features within the subset. The redundancy is calculated as follows:

[0080] ;

[0081] in, represents the first composite sample entropy Redundancy; represents the second composite sample entropy variable; Indicates that the target pressure injury category b belongs to the target pressure injury category to which the second composite sample entropy belongs; represents the first composite sample entropy The target pressure injury category is And the second composite sample entropy The joint probability that the target pressure injury category is b; represents the second composite sample entropy The marginal probability that the target pressure injury category is b.

[0082] Based on the correlation and its corresponding redundancy, a weight score is calculated; the calculation formula of the weight score is:

[0083] .

[0084] The weighted scores are normalized to obtain a weighted weight of the entropy of each composite sample.

[0085] The product of the weighted weight and the corresponding composite sample entropy is used as the weighted sample entropy, and the weighted sample entropy is arranged to obtain the weighted sample entropy feature vector. Finally, the FrCMSE feature vector is converted into a weighted feature vector WFrCMSE:

[0086] ;

[0087] The method of constructing a training sample set based on the sample entropy feature vector to train a support vector machine classifier to obtain a pressure injury staging model includes:

[0088] The training sample set is constructed based on the sample entropy feature vector.

[0089] The radial basis function is used as the kernel function of the support vector machine classifier; the expression of the kernel function is:

[0090] ;

[0091] in, represents the kernel function; Indicates the weighted sample entropy feature vector; Indicates the weighted sample entropy feature vector; exp represents the exponential function; Represents the kernel function parameter, which is used to control the width of the kernel function; ‖*‖ represents the Euclidean norm.

[0092] When training the support vector machine classifier, the optimal hyperplane is found by optimizing the model objective function to obtain the pressure injury staging model; the expression of the model objective function is:

[0093] ;

[0094] Where, Indicates finding the minimum value of the hyperplane normal vector W and the bias term d; W represents the hyperplane normal vector; d represents the bias term; Represents the regularization parameter, which controls the fault tolerance of the classifier. The value will reduce the training error, but may lead to overfitting; q represents the training sample variable; Q represents the total number of training samples; represents the slack variable used to handle classification errors.

[0095] The pressure injury video to be tested is input into the pressure injury staging model to obtain the injury staging result of the pressure injury video to be tested. The new video data is automatically detected using the trained pressure injury detection model and the detection result is output.

[0096] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A pressure injury detection method based on fractional-order composite multi-scale sample entropy, characterized in that: include: Dynamically tracking the region of interest in the pressure injury video based on the adaptive optical flow field to obtain a tracking video of the region of interest; Processing the tracking video by Euler video magnification method to construct a brightness image sequence; Performing fractional differential processing on the brightness image sequence, calculating multi-scale sample entropy eigenvalues, and constructing a sample entropy eigenvector based on the multi-scale sample entropy eigenvalues, including: performing nonlinear enhancement on the brightness image sequence by fractional-order differentiation to obtain fractional-order derivatives at multiple time points, and arranging the fractional-order derivatives to obtain a fractional-order derivative sequence; performing coarse-graining processing on the fractional-order derivative sequence based on a plurality of scale factors to obtain a multi-scale coarse-grained sequence; Performing multi-scale sample entropy calculation on the multi-scale coarse-grained sequence to obtain the sample entropy feature vector includes: calculating the sample entropy of each coarse-grained sequence in the multi-scale coarse-grained sequence; calculating the composite sample entropy under multiple scale factors based on the sample entropy; and arranging the composite sample entropies to obtain the sample entropy feature vector; constructing a training sample set based on the sample entropy feature vector to train a support vector machine classifier to obtain a pressure injury staging model; The pressure injury video to be tested is input into the pressure injury staging model to obtain an injury staging result of the pressure injury video to be tested.

2. The pressure injury detection method based on fractional-order composite multi-scale sample entropy according to claim 1 is characterized in that: The method of dynamically tracking the region of interest of the pressure injury video based on the adaptive optical flow field includes: Determine the image brightness variance based on the degree of illumination change of the video frame image; Determining an image gradient mean based on the motion complexity of the video frame image; Determining an adaptive weight coefficient of an adaptive optical flow field based on the image brightness variance and the image gradient mean; Based on the adaptive weight coefficients, an optimization objective function of the brightness consistency error of the adaptive optical flow field is constructed.

3. The pressure injury detection method based on fractional-order composite multi-scale sample entropy according to claim 2 is characterized in that: The calculation formula of the adaptive weight coefficient is: ； The calculation formula of the optimization objective function of the brightness consistency error is: ； in, represents the adaptive weight coefficient; and Represent the brightness weight coefficient and gradient weight coefficient respectively; represents the image brightness variance of frame images in the pressure injury video; Represents the image gradient mean of the frame image in the pressure injury video; Indicates the brightness value of the pressure injury video; represents the spatial gradient vector of brightness; represents the gradient modulus; represents the value of the optimization objective function of brightness consistency error; i represents the pixel variable in the region of interest; N represents the total number of pixels in the region of interest; Represents the brightness gradient of the frame image in the pressure injury video in the x direction; Represents the brightness gradient of the frame image in the pressure injury video in the y direction; Represents the temporal gradient of the brightness of the frame image in the pressure injury video; Represents the displacement of the i-th pixel variable in the x direction; Represents the displacement of the i-th pixel variable in the y direction; Represents the displacement gradient of the i-th pixel variable in the x direction; Represents the displacement gradient of the i-th pixel variable in the y direction.

4. The pressure injury detection method based on fractional-order composite multi-scale sample entropy according to claim 1 is characterized in that: The nonlinear enhancement of the brightness image sequence by fractional-order differentiation is performed to obtain the calculation formula of the fractional-order derivatives at multiple time points: ； in, Indicates at a point in time The fractional derivative at ; t represents the time variable; Represents the brightness image sequence at time point t; It indicates the value of the calculation formula when the sampling time interval approaches 0 from a direction greater than 0; h indicates the sampling time interval; represents the fractional order; k represents the length of the backtracking time; Represents the lower limit parameter; represents the generalized binomial coefficient; Indicates a time point Brightness image sequence at .

5. The pressure injury detection method based on fractional-order composite multi-scale sample entropy according to claim 1 is characterized in that: The coarse-graining process is performed on the fractional-order derivative sequence based on a plurality of scale factors to obtain a coarse-grained sequence, including: Determining a plurality of scale factors, and dividing the fractional-order derivative sequence into a plurality of non-overlapping time windows of a plurality of time scales according to the time lengths of the plurality of scale factors; Calculate the average value of the fractional derivative in each non-overlapping time window to obtain the coarse-grained data points; Arranging the coarse-grained data points of each scale factor respectively to obtain a coarse-grained sequence; The coarse-grained sequences of multiple scales are arranged to obtain a multi-scale coarse-grained sequence.

6. The pressure injury detection method based on fractional-order composite multi-scale sample entropy according to claim 1 is characterized in that: Calculating the sample entropy of each coarse-grained sequence in the multi-scale coarse-grained sequence respectively, including: Constructing an embedding vector of a first dimension to obtain a first number of first templates; Calculate the distance between the first templates, and use the number of first templates whose distance is less than the distance tolerance as the first number parameter; Constructing an embedding vector of a second dimension to obtain a second number of second templates; Calculating the distance between the second templates, and taking the number of the second templates whose distance is less than the distance tolerance as a second number parameter; The sample entropy is calculated based on the first numerical parameter and the second numerical parameter.

7. The pressure injury detection method based on fractional-order composite multi-scale sample entropy according to claim 1 is characterized in that: The method further includes performing weighted processing on the composite sample entropy under multiple scale factors in the sample entropy feature vector to obtain a weighted sample entropy feature vector, including: Dividing the value range of the composite sample entropy to obtain multiple pressure injury categories; Based on the values ​​of the composite sample entropies, respectively determining target pressure injury categories into which the entropies of the multiple composite samples fall; calculating, based on the pressure injury category and the target pressure injury category, a correlation between the composite sample entropy and a plurality of pressure injury categories; determining redundancy between the composite sample entropies based on the target pressure injury category; Based on the correlation and its corresponding redundancy, a weight score is calculated; Normalizing the weighted scores to obtain a weighted weight of the entropy of each composite sample; The product of the weighted weight and the corresponding composite sample entropy is used as the weighted sample entropy, and the weighted sample entropy is arranged to obtain a weighted sample entropy feature vector.

8. The pressure injury detection method based on fractional-order composite multi-scale sample entropy according to claim 7 is characterized in that: The calculation formula of the correlation is: ； in, represents the first composite sample entropy Correlation with multiple pressure injury categories separately; Indicates the target pressure injury category Belongs to the target pressure injury category to which the first composite sample entropy belongs; Indicates the Scaled composite sample entropy; represents the first composite sample entropy variable; Indicates that pressure injury category c belongs to multiple pressure injury categories ; Represents a set of pressure injury categories, including multiple pressure injury categories; represents the joint probability of the target pressure injury category a and the pressure injury category c of the first composite sample entropy; represents the first composite sample entropy The target pressure injury category is The marginal probability of represents the marginal probability of pressure injury category c; The calculation formula of the redundancy is: ； in, represents the first composite sample entropy Redundancy; represents the second composite sample entropy variable; Indicates that the target pressure injury category b belongs to the target pressure injury category to which the second composite sample entropy belongs; represents the first composite sample entropy The target pressure injury category is And the second composite sample entropy The joint probability that the target pressure injury category is b; represents the second composite sample entropy The marginal probability that the target pressure injury category is b; The calculation formula of the weight score is: 。 9. The pressure injury detection method based on fractional-order composite multi-scale sample entropy according to claim 1 is characterized in that: The step of constructing a training sample set based on the sample entropy feature vector to train a support vector machine classifier to obtain a pressure injury staging model includes: Constructing the training sample set based on the sample entropy feature vector; The radial basis function is used as the kernel function of the support vector machine classifier; the expression of the kernel function is: ； in, represents the kernel function; Indicates the weighted sample entropy feature vector; Indicates the weighted sample entropy feature vector; exp represents the exponential function; represents the kernel function parameter; ‖*‖ represents the Euclidean norm; When training the support vector machine classifier, the optimal hyperplane is found by optimizing the model objective function to obtain the pressure injury staging model; the expression of the model objective function is: ； Where, Indicates finding the minimum value of the hyperplane normal vector W and the bias term d; W represents the hyperplane normal vector; d represents the bias term; represents the regularization parameter; q represents the training sample variable; Q represents the total number of training samples; represents the slack variable.

Images