LC-OCT high-resolution skin image reconstruction method

Abstract

The invention provides an LC-OCT high-resolution skin image reconstruction method. The method comprises the following steps: (1) collecting original interference data of skin tissues by using an LC-OCT system; (2) taking the data acquired in the step (1) as sample data, obtaining frequency interference data by adopting an experimental measurement or theoretical simulation mode, and then constructing depth label data based on a linear relation between frequency and depth; (3) executing a series of preprocessing operations on the data obtained in the step (2); (4) substituting the frequency interference data and the depth label data obtained in the step (3) into a linear regression model for training, and then substituting the preprocessed original interference data into the trained model for prediction output; and (5) obtaining a high-resolution LC-OCT reconstructed image of the skin tissue. Therefore, high-resolution skin tissue image reconstruction can be realized, more skin tissue structure details are provided for dermatologists, and the doctors are helped to diagnose skin diseases more accurately.

Description

Technical Field

[0001] The present invention relates to the technical field of LC-OCT image processing, and in particular to a LC-OCT high-resolution skin image reconstruction method. Background Art

[0002] The full name of OCT in Chinese is optical coherence tomography technology. It is a high-resolution biomedical imaging technology that was born in the 1990s. It can quickly provide high-resolution cross-sectional and three-dimensional imaging. It has developed rapidly due to its advantages such as non-destructive, strong penetration and fast scanning speed, and has received widespread attention and recognition from the society.

[0003] OCT technology is based on low-coherence interferometry and is generally divided into three categories: time-domain OCT (TD-OCT), spectral-domain OCT (SD-OCT), and swept-spectrum OCT (SS-OCT). With the continuous development of science and technology, the application areas of OCT technology have expanded from ophthalmology to other biomedical fields such as dermatology, dentistry, and cardiovascular disease. In addition, it is also involved in materials science and industrial testing. After technological iteration, a variety of variant technologies have also emerged, such as polarization OCT (PS-OCT) and full-field OCT (FF-OCT), and line-field confocal OCT (LC-OCT) is one of them.

[0004] LC-OCT is a technology that combines TD-OCT and confocal microscopy. It measures the echo delay and amplitude of light backscattered from the sample structure through low-coherence interferometry. Different from traditional point scanning, this technology uses line illumination and line detection to acquire multiple single-depth A-Scan signals in parallel. That is, a single longitudinal scan can collect multiple A-Scan signals at different depths for B-Scan imaging. If a scanning galvanometer is used for transverse scanning, a C-Scan signal can be obtained for three-dimensional structural imaging.

[0005] Longitudinal resolution is one of the key parameters of the OCT system, which indicates the ability to resolve structures in the depth direction of the sample. The theoretical longitudinal resolution is determined by the coherence length and can be expressed as follows: Where λ0 represents the central wavelength of the laser light source, n s represents the refractive index of the sample medium, and Δλ represents the full width at half maximum (FWHM) value of the spectral bandwidth of the laser light source.

[0006] As can be seen from the above formula, the theoretical longitudinal resolution is limited by the central wavelength and spectral bandwidth of the light source. Although existing technologies use shorter central wavelengths or wider spectral bandwidths, the high system complexity and bandwidth limitations of specific wavelengths still affect image reconstruction. Summary of the Invention

[0007] The present invention aims to solve one of the technical problems in the related art at least to a certain extent.

[0008] To this end, the purpose of the present invention is to propose a LC-OCT high-resolution skin image reconstruction method. This method is not limited by the light source parameters in the LC-OCT system: the central wavelength of the light source and the spectral bandwidth of the light source. It can achieve high-resolution skin tissue image reconstruction and provide dermatologists with more skin tissue structure details, thereby helping doctors to diagnose skin diseases more accurately.

[0009] To achieve the above objectives, the present invention proposes a LC-OCT high-resolution skin image reconstruction method, comprising the following steps:

[0010] (1) Using the LC-OCT system to collect raw interferometric data of skin tissue;

[0011] (2) Using the data collected in step (1) as sample data, frequency interference data is obtained by experimental measurement or theoretical simulation, and then depth label data is constructed based on the linear relationship between frequency and depth;

[0012] (3) performing a series of preprocessing operations on the data obtained in step (2);

[0013] (4) Substitute the frequency interference data and depth label data obtained in step (3) into the linear regression model for training, and then substitute the pre-processed original interference data into the trained model to predict the output;

[0014] (5) Obtain high-resolution LC-OCT reconstructed images of skin tissue.

[0015] In addition, the LC-OCT high-resolution skin image reconstruction method proposed in the application may also have the following additional technical features:

[0016] Specifically, the step (2) includes the following steps:

[0017] (2.1) Acquisition method: Select the frequency interferometry data based on the acquisition difficulty. If the LC-OCT system is easy to operate during the measurement process, choose the experimental measurement method; otherwise, choose the theoretical simulation method.

[0018] (2.2) Experimental method: Skin tissue is placed in the sample arm of the system. Line illumination and line detection provide lateral scanning. The sample in the sample arm uses a piezoelectric stage to scan the reflector of the reference arm longitudinally from zero optical path difference to obtain interference signals at different depths.

[0019] (2.3) Theoretical simulation method: simulate the relationship between the number of pixels and time delay at different depths, extract a B-Scan signal from the original interferometer signal, first take the average of the column vector of the signal to obtain a row vector, then multiply the row vector by a frequency term, where the frequency term includes the depth variable, and then take the depth variable as a positive integer value to obtain the frequency interferometer data;

[0020] (2.4) According to step (2.3), each time the depth variable value increases, a standard basis vector is obtained, where the standard basis vector is non-zero only in the elements of the corresponding row number. These standard basis vectors are then combined from top to bottom by row to obtain the depth label data.

[0021] Specifically, the step (3) includes the following steps:

[0022] (3.1) Perform four preprocessing steps on the raw interferometer data and the frequency interferometer data obtained through experimental measurement or theoretical simulation: correlation analysis, DC term subtraction, standardization, and principal component analysis (PCA). To avoid data leakage and maintain data consistency, except for the second step of subtracting the DC term, the statistical information obtained during the preprocessing of the frequency interferometer data should be used to perform corresponding preprocessing operations on the raw interferometer data. When using principal component analysis, the square of the prediction residual, PRESS, and F-Value can be calculated to determine the number of principal components to be selected.

[0023] (3.2) Perform two preprocessing operations on the deep label data in sequence: data format conversion and standardization.

[0024] Specifically, the step (4) includes the following steps:

[0025] (4.1) Each column vector of the deep label data is a label. A linear regression model is built by combining each column label data with the preprocessed frequency interference data. After the model is trained, the preprocessed raw interference data is substituted into the model for column-by-column prediction output. There are two linear regression models: a ridge regression model and a model based on the sparrow search algorithm, support vector regression, and partial least squares.

[0026] (4.2) Based on the ridge regression model, the l2 norm is introduced on the basis of the ordinary least squares method. The model training is adjusted by adjusting the regularization parameter. The regularization parameter is set to several initial values. The optimal parameters and the optimal training model are found by cross-validation based on the mean square error. The mean square error is calculated based on the predicted value and the true value.

[0027] (4.3) Based on the sparrow search algorithm + support vector regression + partial least squares model, the estimated value of the light intensity of the preprocessed raw interference data is obtained. Among them, the support vector regression uses a linear kernel, so there are two hyperparameters: the penalty factor C and the insensitive loss function value ε;

[0028] (4.4) The estimated value of the light intensity of the original interference data obtained in step (4.3) is obtained. The reconstructed image has obvious stripe noise, which makes the structure of the skin tissue unclear. A label error model based on partial least squares is introduced to obtain the estimated value of the light intensity of the corrected and preprocessed original interference data.

[0029] Specifically, the step (5) includes the following steps:

[0030] (5.1) Using the corresponding drawing function, the estimated value of the light intensity of the corrected and preprocessed raw interference data obtained in step (4.4) is used to obtain a high-resolution skin LC-OCT reconstructed image.

[0031] Specifically, the step (4.3) includes the following steps:

[0032] (4.3.1) First, set the upper and lower bounds of the two hyperparameters, then set the relevant parameter values ​​of the sparrow search algorithm. Then, the sparrows search within the set hyperparameter range. The position of each sparrow represents the hyperparameter value to be optimized. The mean squared error between the predicted and actual output values ​​is used as the fitness function of the sparrow search. The fitness values ​​of all sparrows are calculated to find the current global optimal position.

[0033] (4.3.2) Update the positions of producers, pickers, and alerters in the sparrow population accordingly, check the individual fitness values ​​after the position update, determine the position of the best individual sparrow, and determine whether the maximum number of iterations is met. If so, output the best sparrow position and use it as the optimized value of parameters C and ε; otherwise, continue to execute the position update rule;

[0034] (4.3.3) The optimized value outputted at the end is used as the optimal parameter value to be substituted into the support vector regression for training and prediction, and the estimated value of the light intensity of the pre-processed raw interference data is obtained.

[0035] Specifically, the step (4.4) includes the following steps:

[0036] (4.4.1) First, the pre-processed frequency interference data is re-substituted into the support vector regression model optimized by the sparrow search algorithm and the predicted label data is predicted. The predicted label data is then subtracted from the deep label data to obtain the label error data.

[0037] (4.4.2) For each column of errors, use the partial least squares model to perform regression;

[0038] (4.4.3) Finally, the original interference data is substituted into the regression model to predict the prediction error value. The error is corrected by taking the difference between the estimated value obtained in step (4.3) and the predicted error value to obtain the estimated value of the light intensity of the corrected and preprocessed original interference data, thereby achieving a more accurate image reconstruction result.

[0039] Compared with the existing technology, the present invention has the following advantages: 1. The present invention constructs a linear regression model based on frequency and depth to achieve LC-OCT image reconstruction, which can improve the longitudinal resolution and obtain high-resolution skin reconstructed images. High-resolution skin image reconstruction will provide dermatologists with higher-resolution LC-OCT images, thereby observing more details of skin tissue structure, helping doctors to more accurately diagnose skin diseases such as basal cell carcinoma and melanoma;

[0040] 2. Apply four methods: correlation analysis, DC term subtraction, standardization, and principal component analysis. Correlation analysis can reduce redundant data and avoid multicollinearity, thereby improving the stability and interpretability of the model; DC term subtraction can eliminate the offset between different data samples; standardization can eliminate dimensional differences between features and reduce computational complexity; principal component analysis can further reduce data dimensionality and improve model performance. In addition, data format conversion of deep label data helps maintain numerical stability during model training.

[0041] 3. The ridge regression model introduces the l2 norm and the corresponding regularization parameter λ into the loss function of ordinary least squares to form an L2 regularization term, which helps to reduce the risk of model overfitting;

[0042] 4. Based on the model of sparrow search algorithm + support vector regression + partial least squares, the sparrow search algorithm is used instead of the traditional K-fold cross-validation method to optimize the hyperparameters C and ε of support vector regression, which helps to reduce computational overhead, experimental cost and model complexity. In addition, partial least squares is used to construct a label error model on the above basis to remove stripe noise, thereby obtaining higher quality image reconstruction results.

[0043] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0045] Figure 1 It is the overall flow chart of the method of the present invention;

[0046] Figure 2 Schematic diagram of the LC-OCT system of the method of the present invention;

[0047] Figure 3 Schematic diagram of the spectral data preprocessing steps of the method of the present invention;

[0048] Figure 4 Schematic diagram of the ridge regression model based on the method of the present invention;

[0049] Figure 5 Flowchart of the SSA+SVR+PLS model of the method of the present invention;

[0050] Figure 6 Schematic diagram of the label error model of the method of the present invention. DETAILED DESCRIPTION

[0051] This application involves multiple scientific and technological fields such as medicine, LC-OCT, and image processing, and covers a wide range. Therefore, unless otherwise defined, the meanings of all technical and scientific terms used in this document are the same as those commonly understood by technicians in the field of this application; the terms described in the text are only used for the purpose of describing specific embodiments and are not meant to limit this application; the terms "including" and "having" used in this application, as well as any variations thereof, are intended to cover non-exclusive meanings, as shown in the claims and the above-mentioned drawings.

[0052] A method for reconstructing high-resolution skin images using LC-OCT will be described below with reference to the accompanying drawings.

[0053] like Figure 1 , an LC-OCT high-resolution skin image reconstruction method according to an embodiment of the present invention comprises the following steps:

[0054] (1) Use the LC-OCT system to collect the original interference data of skin tissue, such as Figure 2As shown in the figure, the system primarily consists of a near-infrared broadband light source, a Michelson interferometer, and a line scan camera. Light from the broadband light source is split into two beams by a beam splitter. One beam enters the reference arm, composed of a microscope objective and a reflector, and returns; the other beam enters the sample arm, composed of a scanning galvanometer, a microscope objective, and the sample, and returns. The two returning beams interfere with each other at the beam splitter and are detected by the line scan camera. A collimator collimates the broadband light, and a cylindrical lens converts the collimated beam from a point source into a line source, providing line illumination. Microscopes in the reference and sample arms focus the beams, and a folding mirror redirects the optical path. In particular, dichroic filters are used as filters to filter light of different wavelengths, adapting to various imaging techniques. The scanning galvanometer performs lateral scanning (in the Y direction). A tubular lens focuses or parallelizes the interfering beams to ensure that the light signal scattered back from the sample can be correctly received by the line scan camera. Finally, a piezoelectric stage scans the entire Michelson interferometer in the longitudinal direction (in the Z direction).

[0055] Therefore, there are two 3D imaging modes for the LC-OCT system: line illumination and line detection, to detect the interference signals of all pixels in the X direction, and the interference data of each pixel is represented as an A-Scan.

[0056] First, if the piezoelectric platform first performs a longitudinal scan to obtain a B-Scan signal for vertical cross-sectional imaging, then the scanning galvanometer performs a transverse scan to obtain a C-Scan signal for three-dimensional structural imaging (horizontal stacking); Second, if the scanning galvanometer first performs a transverse scan to obtain a B-Scan signal for horizontal cross-sectional imaging, then the piezoelectric platform performs a longitudinal scan to obtain a C-Scan signal for three-dimensional structural imaging (vertical stacking). Consistent with traditional OCT imaging methods, the present invention adopts the results of method 1 to improve axial resolution.

[0057] (2) Extract the B-Scan data collected by the system in step (1) as sample data, use experiments or simulations to obtain frequency interference data, and then construct depth label data based on the linear relationship between frequency and depth, which has the following characteristics:

[0058] (2.1) The method chosen to obtain frequency interference data depends on the difficulty of obtaining it. If measuring ex vivo skin tissue, an experimental measurement method can be selected. However, if measuring living skin tissue, the system operation using the experimental measurement method becomes difficult, so a theoretical simulation method can be selected.

[0059] (2.2) The experimental measurement method is to place skin tissue in the sample arm of the system, and line illumination and line detection provide lateral scanning, while the sample in the sample arm uses a piezoelectric platform to scan the reflector of the reference arm longitudinally from zero optical path difference to obtain interference signals at different depths.

[0060] (2.3) The theoretical simulation method is to simulate the relationship between the number of pixels and time delay at different depths. A B-Scan signal is extracted from the original interference signal, denoted as I(τ). Assuming that the signal is an m×n matrix, where m is the different lateral positions and n is the different depth positions, it is divided into n m×1 column vectors according to the matrix rule. The average value of these column vectors is taken. For example, the first column can be expressed as: Among them, j = 1, 2, ..., m, corresponding to different lateral positions in the interference signal.

[0061] The above process can be recursively reduced to: If all the elements calculated by the above formula are Arrange them in order into a 1×m row vector, denoted as Frequency term 1+|A|cos(ω·τ n )(i=1,2,…,n) as the multiplicand and Multiply, where ω = 2πf0, f0 is the Doppler frequency, and f0 = 2V / λ0, V is the scanning speed of the piezoelectric platform, and λ0 is the central wavelength of the light source; assuming that the speed of light is c and the refractive index of the sample is n(z), then Among them, z d is the "depth", and there is a relationship: z d ∝d, where d represents the number of frequency signals (row vectors) in the frequency interference signal. In particular, considering the rules of matrix operations, this multiplication should be element-wise multiplication. For example, when d = 1, a frequency signal (row vector) can be obtained, and its elements can be calculated as:

[0062]

[0063] The above process can be recursively reduced to: Where f = z d , then the above formula can be written as I orig.1,i (2) If d = 2, then f = z d =4, we can also get I orig.2,i (4), so we can conclude that in,[*] a×b Represents a matrix with rows and columns of b, d = 1, 2, ..., m', I orig.d,i (f) is the frequency interference data, and the value of max{d}, i.e., m′, is set according to the specific situation, but it will determine the depth of the reconstructed image.

[0064] The |A| in the frequency term represents the amplitude, which is also set according to the specific situation: there is no obvious difference in the light intensity amplitude between the frequency interference signal and the original interference data. In addition, according to z d ∝d and f = zd , we can get the following relationship: I orig.d,i (f) = I orig.d,i (z d )=I orig.d,i (2·d).

[0065] (2.4) According to step (2.3), I orig.d,i (f) is composed of m′ 1×n row vectors, i.e., frequency spectra, and has a depth of z d = m′, then we construct a corresponding depth label for each frequency spectrum. At this time, each time the depth variable value increases, a standard basis vector is obtained, where the standard basis vector is non-zero only in the element value of the corresponding row number, for example, I orig.d,i The depth row label of (2) is e1=(1,0,…,0) 1×m′ Therefore, these vectors are combined row by row from top to bottom to obtain the deep label data. This process can be recursively as follows: In practice, the input format of a single "label" in linear regression modeling is a column vector. In addition, the identity matrix E has transposition invariance: E T =E, so the above formula can be rewritten as: In particular, the non-zero element value in the depth tag data is not necessarily "1", but The setting method is used to better match the input features.

[0066] (3) Perform a series of preprocessing operations on the data obtained in step (2).

[0067] (3.1) For the original interferometer data and the frequency interferometer data collected experimentally or constructed theoretically, as Figure 3 As shown, four preprocessing operations need to be performed in sequence: correlation analysis, DC term subtraction, standardization and principal component analysis (PCA), assuming that the frequency interference data X, where each row represents a certain depth sampling point at all lateral positions of the sample, and each column represents all depth sampling points at a certain lateral position of the sample.

[0068] For X, the first step is correlation analysis. A Boolean vector h is defined to identify highly correlated features. By adopting the reverse selection method, we retain the features that show low correlation. The process can be expressed as: X (1) =X(:,~h); where h is calculated by calculating the Pearson coefficient of feature i and feature j |R ij |And set the condition range to constrain, where |R ij |It can be obtained by calculating the covariance and standard deviation of the features, which can be expressed as: Among them, cov(X i ,X j) is the covariance of feature i and feature j, and are the standard deviations of feature i and feature j, respectively.

[0069] So a Boolean value can be represented as: Usually, when the correlation is higher than 0.80, redundant information will be introduced, which will further lead to multicollinearity problems. Therefore, the "value" in the above formula can be set to an empirical value of 0.80.

[0070] The second step is to subtract the DC term and calculate the mean of the i-th sample The column vector of X is subtracted, and the process can be expressed as: (2) =X (1) -μ (1) ; The third step is standardization, calculating the mean of the jth feature and standard deviation It can be obtained that the process can be expressed as: X (3) =(X (2) -μ (2) ) / σ (2) .

[0071] The last step is principal component analysis PCA, based on the standardization, (3) Calculate the covariance matrix C and perform eigenvalue decomposition to obtain the eigenvalue μ i and the corresponding eigenvector η i , the process can be expressed as Then descending order (μ1≥μ2≥…≥μ n ) to obtain the eigenvalue diagonal matrix Λ and the eigenvector matrix E, which can be expressed as: Then select the first s principal components, and the corresponding eigenvectors are recorded as E s , put X (3) Projected onto these principal components, the process can be expressed as: (4) =X (3) E s ; Among them, the number of principal components s is determined by the cumulative explained variance ratio γ s The decision, which usually achieves 95% of the ability to retain absolutely no important information, can be expressed as: Among them, a i represents the proportion of variance explained by the i-th principal component, where i is the intermediate variable and i = 1, 2, …, s or i = 1, 2, …, n.

[0072] PRESS is sometimes combined with leave-one-out cross-validation, but this is not suitable for data with a large number of features, so it can be changed to K-fold cross-validation. The F-Value is further obtained based on the PRESS value: Where PRESS(q) is the PRESS value when using the first q principal components for K-fold cross validation, min(PRESS) is the minimum PRESS value among all principal components, and F(q) represents the calculated F-Value.

[0073] When F(q) is close to 1, it indicates that the effect of using q principal components is similar to the effect of using the optimal number of principal components. In the MATLAB environment, calling the "finv" function can achieve the above operation. According to the Haaland and Thomas criterion, the significance level α = 0.25 can be selected.

[0074] In order to avoid data leakage and maintain data consistency, in addition to subtracting the DC term in the second step, the statistical information in the preprocessing process of the frequency interference data should be used to perform corresponding preprocessing operations on the original interference data to obtain the preprocessed original interference data.

[0075] (3.2) Deep label data, denoted as L, in addition to the data format conversion from integer to floating point data, it also needs to be standardized to obtain the pre-processed deep label data L norm .

[0076] (4) Substitute the frequency interference data and depth label data obtained in step (3) into the linear regression model for training, and then substitute the preprocessed original interference data into the trained model to predict the output.

[0077] (4.1)L norm Each column vector of X is a label, (4) and L norm Apply linear regression model training column by column, get the model coefficient after model training, and then calculate the model coefficient according to the To predict, the light intensity prediction values ​​at all lateral positions at the same depth are output column by column, and finally integrated to obtain the B-Scan reconstructed image. Among them, the present invention provides two linear regression models as implementation cases for illustration.

[0078] (4.2) Implementation Case 1: Based on the ridge regression model, such as Figure 4 As shown. Introducing the l2 norm into the framework of ordinary least squares method can get the ridge regression model. Among them, the input layer, linear layer, loss function and output layer constitute the structure of ordinary least squares method. Assume that there are m observations in the input layer, and each observation contains n independent variables x i1 ,x i2 ,…,x in and a corresponding dependent variable y i , then the linear layer has the expression: i =β0+β1x i1 +β2xi2 +…+β n x in +ε i ; where x i1 ,x i2 ,…,x in Also called the predictor variable, y i Also called response variables, β0,β1,β2,…,β n is the regression coefficient, β0 is the intercept term, ε i is the error amount. The above formula has a general expression in matrix form: y = Xβ + ε. The loss function is constructed by minimizing the error term, which can be expressed as: Among them, ||y-Xβ|| 2 It is the square Euclidean norm, which usually represents the residual adjustment sum, so the loss function can also be understood as minimizing the residual sum of squares.

[0079] In the output layer, in order to obtain the estimated value of the linear regression coefficient and predicted values First, simplify L(β): L(β) = (y - Xβ) T (y-Xβ)=…=y T y-2y T Xβ+β T X T Xβ; where L(β) represents the error term, [*] T Represents a transpose operation on a matrix.

[0080] Secondly, the solution is obtained by gradient calculation and inverse operation Among them, solving The process can be summarized as a closed-form solution, (X T X) -1 Is the matrix (X T X), the inverse matrix of T is the transpose of matrix X, It means partial derivative.

[0081] Similarly, we can also derive the loss function of the ridge regression model, the estimated and predicted values ​​of the linear regression coefficients, which can be expressed as follows:

[0082] Among them, ||β|| 2 is the l2 norm, λ is the regularization parameter, λ||β|| 2 This constitutes the L2 regularization term, where λ controls the weight and is usually set to 10 x ,x=0,±1,±2,…,I is the identity matrix, where L norm,j,j=1,2,…,m′ represents the jth depth label data after preprocessing, The preprocessed raw interference data Substitute the predicted value of the jth dimension in the trained model. The λI in the closed-form solution of is positive definite and reversible.

[0083] Among them, the regularization parameter is set to several initial values: λ = 10 x ,x=0,±1,±2,±3,±4,±5, the training data is divided into 4-fold training sets and 1-fold validation sets through 5-fold cross validation, where the validation sets are alternately formed into 5 subsets, and then the optimal parameters are found and the optimal training model is determined based on the mean square error, where the mean square error evaluation is calculated based on the predicted value and the true value:

[0084] (4.3) Implementation Case 2: Model based on sparrow search algorithm + support vector regression + partial least squares, X (4) and L norm A column of is used as a training dataset for a single support vector regression (SVR) model: in is the input feature vector, is the target value, n is the number of samples, and p is the feature dimension. Then the linear regression model constructed by the SVR in the high-dimensional space can be expressed as: f(x) = w T x+b; where is the weight vector, is the bias term.

[0085] In practice, in order to ensure the feasibility of the constraints, a slack variable ξ is introduced within a certain error range. i and After that, the optimization problem can be expressed as: Among them, C>0 is the penalty factor, ε is the insensitive loss function, and defines the tolerance range of error.

[0086] Use the Lagrange multiplier method to solve, that is, construct the Lagrange function and calculate w, b, ξ i , Take the derivative and substitute w into the original Lagrangian function to get the dual problem:

[0087] Among them, α i , is the Lagrange multiplier. After the dual problem is optimized, the regression function used for prediction can be expressed as: Here, b is determined by the Karush-Kuhn-Tucker (KKT) condition for solving the dual problem.

[0088] The SVR model in the present invention uses a linear kernel function, and also has two hyperparameters: C and ε, where the default value of C is 1 and the default value of ε is 0.1.

[0089] To find the best hyperparameters in the SVR model, cross-validation methods are often used, such as K-fold cross-validation and leave-one-out cross-validation. Cross-validation requires evaluating model performance multiple times, which is computationally expensive for many folds, large datasets, or complex models. In particular, this method also requires setting several parameter values ​​for selection. Due to the differences between samples, it will be difficult to set uniform and accurate parameter values, thereby increasing model complexity and experimental costs.

[0090] Influenced by the advantages of metaheuristic algorithms in optimization, SSA replaces 5-fold cross validation to optimize the hyperparameters in the SVR model, that is, searching and selecting the best penalty factor C and insensitive loss function ε within a certain range to improve the prediction accuracy of the model.

[0091] The Sparrow Search Algorithm (SSA) is an effective and novel swarm optimization algorithm inspired by the foraging behavior and anti-intrusion strategies of sparrows. It divides the sparrow population into producers and foragers according to their energy reserve levels, and conducts food-searching activities within the area. The energy reserve level depends on the assessment of individual fitness values. A certain proportion of sparrows in the sparrow population act as early warning systems. The positions of individual sparrows are continuously updated.

[0092] Assume that the position of the sparrow is X i,j , where i = 1, 2, ..., N, j = 1, 2, ..., D, and N and D represent the number of sparrows and the dimension of the variable to be optimized, respectively. Therefore, the position of the sparrow can be expressed in a matrix and recorded as X. Then, the fitness value of the sparrow can be expressed as follows: The value of each row represents the fitness value of an individual. Producers with higher fitness values ​​have priority in obtaining food during the search process, have a larger search range than pickers, and provide foraging areas or directions for pickers. During each iteration, the producer's position update rule is as follows: Among them, iter max Indicates the maximum number of iterations, t indicates the number of current iterations and t=1,2,…,iter max , represents the value of the i-th sparrow in the j-th dimension when the number of iterations is t, the random number α∈(0,1], Q is a random number that obeys the normal distribution, L is a row vector of all 1s with D columns, R2 and ST represent the alarm value and safety threshold respectively, and R2∈[0,1], ST∈[0.5,1].

[0093] Some pickers will frequently monitor producers and try to snatch food. If they succeed, they get the food. Otherwise, they continue to scout. The rules for updating the position of the pickers are as follows: Among them, X p represents the optimal position of the producer, and X w is the current global worst position. In this case, the picker is most likely to starve to death. A is a row vector with D columns and ±1 random elements, and A + =A T (AA T ) -1 .

[0094] Sparrows that act as early warning systems can sense danger. They account for 10% to 20% of the total number of sparrows, and their initial positions are randomly generated. The early warning system position update rules are as follows: Among them, the step size control parameter β~N(0,1), the random number K∈[-1,1], ∈ is the smallest constant, X best is the current global best position, f i is the current sparrow’s fitness value, f g and f w are the current global best and worst fitness values ​​respectively. In addition, the fitness function is also evaluated using mean square error.

[0095] like Figure 5 As shown in Figure 2, a flowchart of SSA optimizing SVR hyperparameters is given.

[0096] (4.3.1) First, set the upper and lower bounds of the two hyperparameters C and ε, respectively. The default values ​​allow for the following ranges: C ∈ [0.1, 10] and ε ∈ [0.01, 1]. Then, set the relevant parameters of the sparrow search algorithm: a producer ratio of 20%, a picker ratio of 80%, a warning ratio of 20%, a population size of 30–50, and a maximum number of iterations of 100–500, which are suitable for medium-sized problems. Considering computational resources and the introduction of the label error model, the values ​​of these parameters can be appropriately reduced to balance solution accuracy and computational efficiency. The dimension of the variable to be optimized, D, is set to 2. The sparrows then search within the set hyperparameter range, with each sparrow's position representing the hyperparameter value to be optimized. The mean squared error between the predicted and actual output values ​​serves as the sparrow search fitness function, and the fitness values ​​of all sparrows are calculated to find the current global optimal position.

[0097] (4.3.2) Then, according to the position update rules of producers, pickers and early warning in the sparrow population, corresponding updates are made, the individual fitness values ​​after the position update are checked, the position of the best individual sparrow is determined, and it is judged whether the maximum number of iterations is met. If so, the best sparrow position is output and used as the optimized value of parameters C and ε; otherwise, the position update rules are continued.

[0098] (4.3.3) The optimized value outputted at the end is used as the optimal parameter value to be substituted into the support vector regression for training and prediction to obtain the estimated value of the light intensity of the pre-processed raw interference data.

[0099] (4.4) The estimated value of the light intensity of the original interference data obtained in step (4.3) is that the reconstructed image has obvious stripe noise, which makes the structure of the skin tissue unclear. For this reason, a label error model based on partial least squares (PLS) is introduced, as shown in Figure 6 shown.

[0100] (4.4.1) First, the pre-processed frequency interference data is re-substituted into the SVR model after SSA optimization and the predicted label data is predicted. The predicted label data is then subtracted from the deep label data to obtain the label error data. This process can be expressed as: in, is the predicted label matrix, Y train is the deep label matrix, and ΔY is the label error matrix.

[0101] (4.4.2) For each column of errors, PLS is used for regression, where the regression process of the i-th column is as follows: ΔY i =X·β i +ε; where ΔY i is the error vector in column i, X is the frequency interference data, β i is the regression coefficient, ε is the error term, and maximizes X and ΔY i The covariance of the regression coefficient can be estimated (including the intercept term).

[0102] (4.4.3) Finally, the original interference data is substituted into the regression model to obtain the prediction error value. The error is then corrected by subtracting the estimated value obtained in step (4.3) from the prediction error value to obtain a more accurate image reconstruction result. This process can be expressed as: Among them, X test is the original interference data, It's X test Substitute the SSA+SVR model to predict the output result, is the correction value, It's X test Substitute the error value of the PLS model prediction output.

[0103] (5) Obtain high-resolution LC-OCT reconstructed images of skin tissue.

[0104] (5.1) The obtained value in step (4.4) Use the corresponding drawing function to obtain the LC-OCT high-resolution reconstructed image of the skin tissue, for example, the "imshow" function can be used in the MATLAB environment.

[0105] Overall, the present invention can achieve LC-OCT image reconstruction using a medium-configuration computer, which is used to improve axial resolution and obtain high-resolution skin reconstructed images. It also provides two different linear regression model methods, which help doctors accurately distinguish the structural details of skin tissue and promote the development of the field of dermatological diseases.

[0106] In the embodiment of the present invention, all or part of the steps may be performed by operating the relevant hardware through program instructions. These programs may be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc.

[0107] Obviously, the embodiments described above only constitute part of the implementation scheme of the present application, rather than all the implementation schemes. The preferred embodiments of the present application are shown in the accompanying drawings, but are not limited thereto, and the patent scope of the present application is not limited to the contents shown in the accompanying drawings. The present application can be implemented in a variety of different forms. The purpose of providing these embodiments is to more fully illustrate the contents of the present application. Although the aforementioned embodiments are described in detail, it is still possible for those skilled in the art to modify the specific implementation methods therein, or to make equivalent replacements for some technical features. Any equivalent structure covered by the contents of the description and drawings of this application, directly or indirectly applied in other related technical fields, should be included in the scope of patent protection of this application.

[0108] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are exemplary and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and deform the above embodiments within the scope of the present invention.

Claims

1. A LC-OCT high-resolution skin image reconstruction method, characterized in that: The following steps are involved: (1) Using the LC-OCT system to collect raw interferometric data of skin tissue; (2) Using the data collected in step (1) as sample data, frequency interference data is obtained by experimental measurement or theoretical simulation, and then depth label data is constructed based on the linear relationship between frequency and depth; (3) performing a series of preprocessing operations on the data obtained in step (2); (4) Substitute the frequency interference data and depth label data obtained in step (3) into the linear regression model for training, and then substitute the pre-processed original interference data into the trained model to predict the output; (5) Obtain high-resolution LC-OCT reconstructed images of skin tissue.

2. The LC-OCT high-resolution skin image reconstruction method according to claim 1, characterized in that: The step (2) comprises the following steps: (2.1) Acquisition method: Select the frequency interferometry data based on the acquisition difficulty. If the LC-OCT system is easy to operate during the measurement process, choose the experimental measurement method; otherwise, choose the theoretical simulation method. (2.2) Experimental method: Skin tissue is placed in the sample arm of the system. Line illumination and line detection provide lateral scanning. The sample in the sample arm uses a piezoelectric stage to scan the reflector of the reference arm longitudinally from zero optical path difference to obtain interference signals at different depths. (2.3) Theoretical simulation method: simulate the relationship between the number of pixels and time delay at different depths, extract a B-Scan signal from the original interferometer signal, first take the average of the column vector of the signal to obtain a row vector, then multiply the row vector by a frequency term, where the frequency term includes the depth variable, and then take the depth variable as a positive integer value to obtain the frequency interferometer data; (2.4) According to step (2.3), each time the depth variable value increases, a standard basis vector is obtained, where the standard basis vector is non-zero only in the elements of the corresponding row number. These standard basis vectors are then combined from top to bottom by row to obtain the depth label data.

3. The LC-OCT high-resolution skin image reconstruction method according to claim 1, characterized in that: The step (3) comprises the following steps: (3.1) Perform four preprocessing steps on the raw interferometer data and the frequency interferometer data obtained through experimental measurement or theoretical simulation: correlation analysis, DC term subtraction, standardization, and principal component analysis (PCA). To avoid data leakage and maintain data consistency, except for the second step of subtracting the DC term, the statistical information obtained during the preprocessing of the frequency interferometer data should be used to perform corresponding preprocessing operations on the raw interferometer data. When using principal component analysis, the square of the prediction residual, PRESS, and F-Value can be calculated to determine the number of principal components to be selected. (3.2) Perform two preprocessing operations on the deep label data in sequence: data format conversion and standardization.

4. The LC-OCT high-resolution skin image reconstruction method according to claim 1, characterized in that: The step (4) comprises the following steps: (4.1) Each column vector of the deep label data is a label. A linear regression model is built by combining each column label data with the preprocessed frequency interference data. After the model is trained, the preprocessed raw interference data is substituted into the model for column-by-column prediction output. There are two linear regression models: a ridge regression model and a model based on the sparrow search algorithm, support vector regression, and partial least squares. (4.2) Based on the ridge regression model, the l2 norm is introduced on the basis of the ordinary least squares method. The model training is adjusted by adjusting the regularization parameter. The regularization parameter is set to several initial values. The optimal parameters and the optimal training model are found by cross-validation based on the mean square error. The mean square error is calculated based on the predicted value and the true value. (4.3) Based on the sparrow search algorithm + support vector regression + partial least squares model, the estimated value of the light intensity of the preprocessed raw interference data is obtained. Among them, the support vector regression uses a linear kernel, so there are two hyperparameters: the penalty factor C and the insensitive loss function value ε; (4.4) The estimated value of the light intensity of the original interference data obtained in step (4.3) is obtained. The reconstructed image has obvious stripe noise, which makes the structure of the skin tissue unclear. A label error model based on partial least squares is introduced to obtain the estimated value of the light intensity of the corrected and preprocessed original interference data.

5. The LC-OCT high-resolution skin image reconstruction method according to claim 4, characterized in that: The step (5) comprises the following steps: (5.1) Using the corresponding drawing function, the estimated value of the light intensity of the corrected and preprocessed raw interference data obtained in step (4.4) is used to obtain a high-resolution skin LC-OCT reconstructed image.

6. The LC-OCT high-resolution skin image reconstruction method according to claim 4, characterized in that: The step (4.3) comprises the following steps: (4.3.1) First, set the upper and lower bounds of the two hyperparameters, then set the relevant parameter values ​​of the sparrow search algorithm. Then, the sparrows search within the set hyperparameter range. The position of each sparrow represents the hyperparameter value to be optimized. The mean squared error between the predicted and actual output values ​​is used as the fitness function of the sparrow search. The fitness values ​​of all sparrows are calculated to find the current global optimal position. (4.3.2) Update the positions of producers, pickers, and alerters in the sparrow population accordingly, check the individual fitness values ​​after the position update, determine the position of the best individual sparrow, and determine whether the maximum number of iterations is met. If so, output the best sparrow position and use it as the optimized value of parameters C and ε; otherwise, continue to execute the position update rule; (4.3.3) The optimized value outputted at the end is used as the optimal parameter value to be substituted into the support vector regression for training and prediction, and the estimated value of the light intensity of the pre-processed raw interference data is obtained.

7. The LC-OCT high-resolution skin image reconstruction method according to claim 4, characterized in that: The step (4.4) comprises the following steps: (4.4.1) First, the pre-processed frequency interference data is re-substituted into the support vector regression model optimized by the sparrow search algorithm and the predicted label data is predicted. The predicted label data is then subtracted from the deep label data to obtain the label error data. (4.4.2) For each column of errors, use the partial least squares model to perform regression; (4.4.3) Finally, the original interference data is substituted into the regression model to predict the prediction error value. The error is corrected by taking the difference between the estimated value obtained in step (4.3) and the predicted error value to obtain the estimated value of the light intensity of the corrected and preprocessed original interference data, thereby achieving a more accurate image reconstruction result.

Images