Micromedical hyperspectral image dimension reduction method
By introducing Laplace weighted graph embedding and band weighting mechanisms in the tensor dimensionality reduction method, the problems of low robustness of existing methods and neglecting information amounts are solved, and more efficient dimensionality reduction and classification performance improvements in hyperspectral image are achieved.
Patent Information
- Application Number
- CN202411832616.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-12
- Publication Date
- 2025-04-11
AI Technical Summary
The existing tensor-based dimensionality reduction method is not very robust when processing hyperspectral images, and ignores the differences in the amount of information in different samples, resulting in poor classification performance.
A weighted graph embedding method based on Laplace is proposed, which enhances the robustness of the dimensionality reduction method through data preprocessing, band weighting, Laplace graph embedding and tensor dimensionality reduction modules, and improves discrimination through weighting mechanisms.
It effectively reduces the data dimension, retains key information, and enhances classification performance, especially in the diagnosis of gallbladder cancer.
Smart Images
Figure CN119810582A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of dimensionality reduction of medical microscopic hyperspectral images, and more specifically, to a dimensionality reduction method for embedding microscopic hyperspectral images. Background Art
[0002] Hyperspectral imaging technology combines imaging technology and spectral technology to detect the spectral information and image data of the target. Different spectral information represents different physical and chemical structures inside the sample, and different image data reflects different external features and textures of the sample. Microscopic hyperspectral imaging technology combines the characteristics of spectral analysis and optical imaging, and can simultaneously obtain the physical and chemical properties of biological tissue samples, gradually becoming an auxiliary diagnostic tool for intelligent medicine. However, the high-dimensional characteristics of hyperspectral images will lead to poor performance of the classifier, and the increase in the number of bands will cause the data to contain a large amount of redundant information, which is easy to cause the dimensionality disaster and Hughes phenomenon. [3] With the emergence of , we need to perform dimensionality reduction (DR) on hyperspectral image data. This technology removes redundant information by extracting low-dimensional features of information, reducing the feature dimension without losing useful information, and has become an important direction of data processing.
[0003] Hyperspectral image dimensionality reduction is an important data processing technology. It mainly reduces the dimension of data through feature extraction and feature selection, so as to more effectively analyze and utilize hyperspectral information. First of all, feature selection is an important means of hyperspectral image dimensionality reduction. The purpose of feature selection is to select the most representative and informative features from the original feature set to retain key information. The bands in spectral data usually have correlations, and feature selection can help eliminate unnecessary redundant features and improve the generalization ability of the model. Feature extraction is also one of the key steps in hyperspectral image dimensionality reduction. In this process, the original high-dimensional data is mapped to a new feature space through mathematical transformation methods. This method effectively retains the main features in the original data, reduces redundant information, and improves computational efficiency. In general, hyperspectral image dimensionality reduction can effectively retain key information while reducing data dimensions through feature extraction and feature selection, providing more refined data for subsequent analysis and application.
[0004] Hyperspectral images are three-dimensional data that are "graph-spectrum integrated". Traditional vector-based dimensionality reduction methods destroy the spatial structure of hyperspectral images, so a tensor-based dimensionality reduction method is proposed. Tensor local preserving projection is an extended form of local preserving projection in tensor space. In tensor analysis, existing methods ignore the fact that the amount of information contained in samples is different, all samples are treated equally, and the two important issues of enhancing discrimination ability and utilizing inherent structural information are considered separately, which still has great room for development. Summary of the invention
[0005] Purpose of the invention: The purpose of this application is to address the shortcomings of existing tensor-based dimensionality reduction methods and propose a Laplace-based weighted graph embedding method to address the defect of low robustness of the original dimensionality reduction method and enhance the discriminability by weighting the spectral bands.
[0006] Technical solution: To achieve the purpose of the present invention, the technical solution adopted by the present invention is:
[0007] A microscopic hyperspectral image dimensionality reduction method, characterized by: data preprocessing, band weighting module, Laplacian graph embedding module, tensor dimensionality reduction module;
[0008] The preprocessing module mainly performs normalization processing on the original spatial data, that is, standardizes the data so that the data is distributed between 0 and 1.
[0009] The band weighting module calculates the entropy values of all bands. The larger the entropy value, the more information it contains. Different weights are assigned according to the amount of information contained in the band. The preprocessed data is then divided into pixel-patches, and then randomly divided into test sets and training sets.
[0010] The Laplace graph embedding module adds a consensus matrix to obtain the weight matrix for vector dimensionality reduction. The objective function based on sparseness and low rank is constructed, which fully considers the low rank and sparsity of the matrix, and adds a popular regularization term to prevent overfitting. Then, the alternating multiplier method is used to solve the objective function and obtain the target matrix.
[0011] The tensor dimension reduction module is used to obtain a projection matrix from a high-dimensional space to a low-dimensional space. The two spatial dimensions and one spectral dimension are sequentially reduced in combination with the weight matrix of the Laplace cooperative representation to obtain projection matrices in three directions.
[0012] Then, the projection matrix is used to reduce the dimension of the training set and the test set, and the support vector machine is used to classify the samples after dimension reduction. By selecting different parameter values of β, γ, α, window size w, and dimension reduction number b, the classification results are compared to obtain the highest value of sample classification accuracy.
[0013] The specific method of the present invention comprises the following steps:
[0014] Step 1: Medical hyperspectral data is usually regarded as a third-order tensor X∈R l1xl2xl3 , l1, 12 are the length and width of the three-dimensional cube, and I3 represents the total number of dimensions of the spectrum. According to the pixel category, a fourth-order tensor X is defined. (c) ∈R n1xn2xdxI4It is used to represent the collection of pixel blocks containing the same category c.
[0015] Step 2: Calculate the entropy value of each band, and then assign different weights to the spectral bands according to the different entropy values.
[0016] Step 3: Divide the hyperspectral dataset X into pixel-patches, and then randomly divide the pixel-patches into training sets and test sets to facilitate subsequent dimensionality reduction.
[0017] Step 4: Solve each type of weight matrix W according to the objective function c , merged to get the final weight matrix W,
[0018]
[0019] The objective function is:
[0020]
[0021] stdiag(W (c) )=0
[0022] Step 5: Combine the weight matrix to construct the tensor local preservation projection objective function:
[0023]
[0024] Step 6: Set the dimension of the data after dimensionality reduction to {1, 1, d}, and use the mapping matrix U and dimensionality reduction projection formula Y obtained in step 5 i =X i ×1U1×2U2…× n U n , find the sample data y in the low-dimensional space i .
[0025] Step 7: Use support vector machine to classify the samples after dimensionality reduction. By selecting different parameter values of β, γ, α, window size w, and dimensionality reduction number b, the classification results are compared to obtain the highest value of sample classification accuracy.
[0026] On the other hand, the present application discloses an application method for dimensionality reduction of medical microscopic hyperspectral images using the above model, and the specific steps are as follows:
[0027] Step 1: normalize the collected medical microscopic hyperspectral images, i.e., the preprocessing step.
[0028] Step 2: Medical hyperspectral data is usually regarded as a third-order tensor X∈R l1xl2xl3, l1, l2 are the length and width of the three-dimensional cube, and I3 represents the total number of dimensions of the spectrum. Then we calculate the weight of each band separately to perform band weighting. According to the pixel point category, we define a fourth-order tensor X (c) ∈R n1xn2xdxI4 It is used to represent the collection of pixel blocks containing the same category c.
[0029] Step 3: Divide the hyperspectral dataset X into pixel-patches, and then randomly divide the pixel-patches into training sets and test sets to facilitate subsequent dimensionality reduction.
[0030] Step 4: Solve each type of weight matrix W according to the objective function c , merged to get the final weight matrix W,
[0031]
[0032] The objective function is:
[0033]
[0034] stdiag(W (c) )=0
[0035] Step 5: Combine the weight matrix to construct the tensor local preservation projection objective function:
[0036]
[0037] Step 6: Set the dimension of the data after dimensionality reduction to {1, 1, d}, and use the mapping matrix U and dimensionality reduction projection formula Y obtained in step 5 i =X i ×1U1×2U2…× n U n , find the sample data yi in the low-dimensional space.
[0038] Step 7, use support vector machine to classify the samples after dimension reduction, compare the classification results by selecting different parameter values of β, γ, α, window size w, and dimension reduction number b, and obtain the highest value of sample classification accuracy. Beneficial effect: The present invention proposes a new graph embedding method, and uses SVM to analyze the hyperspectral image after dimension reduction. In this method, the medical hyperspectral data is directly input as a third-order vector, which reduces the loss of data and retains the spatial structure information between the data. The spectral bands are weighted before dimension reduction. Through a global attention mechanism, the attention is focused on the bands with more important spectral information. Then, the low rank and sparsity of the matrix are considered, and a popular regularization term is added to prevent the occurrence of overfitting. Then a consensus matrix is constructed to maintain the consistency between the projection matrix and the low rank matrix, better capture the local and global structural information of the data, and enhance robustness. It provides a new technology for the diagnosis of gallbladder cancer, which has research value and significance for doctors to make medical diagnosis of gallbladder cancer. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] Figure 1 The single-band image and true value image of bile duct cancer microscopic hyperspectral image numbered 032236-20x-roi54;
[0040] Figure 2 It is a flowchart for solving the objective function;
[0041] Figure 3 is the overall flow chart of this method;
[0042] Figure 4 This is the classification result diagram of 032236-20x-roi54 obtained by using the method of the present application;
[0043] Table 1 shows the data classification accuracy of the 032236-20x-roi54 image after dimensionality reduction using different dimensionality reduction methods. DETAILED DESCRIPTION
[0044] This paper proposes a new tensor Laplacian graph embedding method, which adds a consensus low-rank matrix when constructing the Laplacian graph to maintain the consistency between the projection matrix and the low-rank matrix. At the same time, considering that different spectral bands contain different amounts of information, a global attention weighting mechanism is used to focus on bands with more important spectral information to improve classification performance.
[0045] In this embodiment, the specific microscopic hyperspectral image dimensionality reduction method includes the following steps:
[0046] Step 1: Perform simple preprocessing on the collected medical microscopic hyperspectral image, that is, normalize the hyperspectral image.
[0047]
[0048] Step 2: Medical hyperspectral data is usually regarded as a third-order tensor X∈R l1xl2xl3 , l1, l2 are the length and width of the three-dimensional cube, and I3 represents the total number of dimensions of the spectrum. Calculate the weight of each band for band weighting. Classify according to the pixel category and define a fourth-order tensor X (c) ∈R n1xn2xdxI4 It is used to represent the collection of pixel blocks containing the same category c. The weighted formula is as follows:
[0049]
[0050] mean(ρ (i，：) ) is the average value of the band.
[0051] Step 3: Divide the hyperspectral dataset X into pixel-patches, randomly divide the pixel-patches into training sets and test sets, and classify them according to the category label c to facilitate subsequent dimensionality reduction.
[0052] Step 4: Solve each type of weight matrix W according to the objective function c , merged to get the final weight matrix W,
[0053]
[0054] The objective function is:
[0055]
[0056] stdiag(W (c) )=0
[0057] Step 5: Combine the weight matrix to construct the tensor local preservation projection objective function:
[0058]
[0059] Step 6: Set the dimension of the data after dimensionality reduction to {1, 1, d}, and use the mapping matrix U and dimensionality reduction projection formula Y obtained in step 5 i =X i ×1U1×2U2…× n U n , find the sample data y in the low-dimensional space i .
[0060] Step 7: Use support vector machine to classify the samples after dimensionality reduction. By selecting different parameter values of β, γ, α, window size w, and dimensionality reduction number b, the classification results are compared to obtain the highest value of sample classification accuracy.
[0061] What needs to be explained in this example is:
[0062] This paper proposes a new graph embedding method, which performs spectral band weighting before dimensionality reduction, focuses on the bands with more important spectral information through a global attention mechanism, and then constructs a consensus matrix to maintain the consistency between the projection matrix and the low-rank matrix, better captures the local and global structural information of the data, and enhances robustness. Then, the low rank and sparsity of the matrix are considered, and a popular regularization term is added to prevent overfitting.
[0063] In tensor analysis, existing methods ignore the fact that the amount of information contained in samples is different. All samples are treated equally, and the two important issues of enhancing discriminative ability and utilizing inherent structural information are considered separately. Moreover, regarding the selection of sample neighborhoods for graph embedding, the selection of neighborhoods needs to rely on pre-set parameters. This process is easily affected by noise, and the obtained neighborhood graph is not robust. Therefore, this paper constructs an integrated framework to solve these problems. A global attention spectrum mechanism is introduced before dimensionality reduction, so that the bands containing large amounts of information have greater weights and play a greater role in the subsequent dimensionality reduction process. A consensus matrix is also added to reveal the relationship between the vector space and the subspace, thereby improving robustness.
[0064] Simulation experiment
[0065] In order to illustrate the effectiveness of the tensor-based microscopic hyperspectral image dimensionality reduction method, the 032236-20x-roi54 cholangiocarcinoma microscopic hyperspectral image was processed and compared with the results of LPP, TLPP, DTME, and DGLPGE. The size of the 032236-20x-roi5 cholangiocarcinoma microscopic hyperspectral image is 1024×1280×60, where 60 is the number of bands. The verification results are based on the python language, and 1% of the samples are randomly selected as the training set. During the training process, γ, β, α are set to 0.1, 0.1, 1, and w is set to 11.
[0066] Experimental results analysis
[0067] Table 1 shows the classification accuracy of different methods when the spectral dimension is reduced to 20. As can be seen from Table 1, the classification accuracy of the present invention is significantly better than that of other methods. It can be seen that the method proposed in the present invention can well extract the features of high-dimensional space and retain them in low-dimensional space, reducing redundancy.
Claims
1. A method for dimensionality reduction of microscopic hyperspectral images, characterized in that: Data preprocessing, band weighting module, Laplacian graph embedding module, tensor dimension reduction module; The preprocessing module mainly performs normalization processing on the original spatial data, that is, standardizes the data so that the data is distributed between 0 and 1. The band weighting module calculates the entropy values of all bands. The larger the entropy value, the more information it contains. Different weights are assigned according to the amount of information contained in the band. The preprocessed data is then divided into pixel-patches, and then randomly divided into test sets and training sets. The Laplace graph embedding module adds a consensus matrix to obtain the weight matrix for vector dimensionality reduction. The objective function based on sparseness and low rank is constructed, which fully considers the low rank and sparsity of the matrix, and adds a popular regularization term to prevent overfitting. Then, the alternating multiplier method is used to solve the objective function and obtain the target matrix. The tensor dimension reduction module is used to obtain a projection matrix from a high-dimensional space to a low-dimensional space. The two spatial dimensions and one spectral dimension are sequentially reduced in combination with the weight matrix of the Laplace cooperative representation to obtain projection matrices in three directions.
2. According to the method for dimensionality reduction of microscopic hyperspectral images in claim 1, it is characterized in that all bands are weighted to different degrees before data dimensionality reduction, so that more attention is paid to important bands during the dimensionality reduction process, data redundancy is reduced, and the efficiency of dimensionality reduction is improved. The calculation formula of band weighting is as follows: x′ band,i =x band,i ×ω band,i ,i=1,2,...,d in, ω is the weight, which is determined by the size of information entropy. We define the calculation formula of information entropy of medical hyperspectral images as: in is the probability that the pixel gray value is g, and n(g) is the number of pixels with gray value g. Then define the amount of information in each band as: where δ band,i ,μ band,i are the standard deviation and mean of the ith band respectively. The correlation coefficient between bands is defined as: Cov(X band,i ,X band,j ) is the covariance between band i and band j, Var(X band,i ), Var(X band,j ) are the variances of bands i and j, respectively. In order to make the bands with large information and small band correlation play a greater role, we assign them a larger weight. The final weight function is defined as:
3. A method for dimensionality reduction of microscopic hyperspectral images according to claim 1, characterized in that A consensus low-rank matrix is added to solve the problem, which makes full use of the relationship between the vector space and the subspace, maintains the consistency between the projection matrix and the low-rank matrix, and enhances the robustness. First, we construct the objective function as: β, γ are hyperparameters for adjusting balance, Z is defined as the Laplace affine matrix of A, and the rows and columns of the matrix A are defined as: is p，c The kth nearest neighbor of (c) The weight matrix W required for each category * is the consensus matrix, ω is the weight matrix, ω ij is the Euclidean distance between the i-th pixel block and the j-th pixel block, defined as t is a hyperparameter, is the frobenius norm, ||·|| * is a sparse matrix solution, ||·||1 is the l1 norm, and the ×4 operation is defined as a 4-module product operation.
4. A supervised dimensionality reduction method for microscopic hyperspectral images, characterized in that: The method comprises the following steps Step 1: Preprocess the collected medical microscopic hyperspectral images, that is, normalize the hyperspectral images. Step 2: Medical hyperspectral data is usually regarded as a third-order tensor X∈R 11x12x13 , l1, l2 are the length and width of the three-dimensional cube, and I3 represents the total number of dimensions of the spectrum. Calculate the weight of each band for band weighting. Classify according to the pixel category and define a fourth-order tensor X(c)∈R n1xn2xdxI4 It is used to represent the collection of pixel blocks containing the same category c. Step 3: Divide the hyperspectral dataset X into pixel-patches, and then randomly divide the pixel-patches into training sets and test sets to facilitate subsequent dimensionality reduction. Step 4: Solve each type of weight matrix W according to the objective function c , merged to get the final weight matrix W, The objective function is: Step 5: Combine the weight matrix to construct the tensor local preservation projection objective function: Step 6: Set the dimension of the data after dimensionality reduction to {1, 1, d}, and use the mapping matrix U and dimensionality reduction projection formula Y obtained in step 5 i =X i ×1U1×2U2…× n U n , find the sample data y in the low-dimensional space i . Step 7: Use support vector machine to classify the samples after dimensionality reduction. By selecting different parameter values of β, γ, α, window size w, and dimensionality reduction number b, the classification results are compared to obtain the highest value of sample classification accuracy.