A feature extraction and classification method for renal tumor MRI tomography images

Abstract

A method for feature extraction and classification of MRI tomographic images of kidney tumors belongs to the field of image recognition technology. The present invention uses MRI tomographic images of actual kidney tumor patients as experimental data and the different Ki67 expression results of the patients as classification criteria, and classifies the images by establishing feature vectors. When extracting features, the characteristics of the MRI tomographic images of kidney tumors are taken into consideration, and the images are fitted using a Gaussian mixture model to obtain mathematical features representing the distribution of the images. The features are then screened to obtain representative image features. A support vector machine is used to classify the features of different images, and a connection is established between the MRI images of kidney tumors and the expression of Ki67, thereby obtaining more accurate image recognition and classification results. The present invention has physical interpretability and can assist doctors in clinical medicine in determining the pathological stage and malignant tendency of patients' kidney tumors.

Description

Technical Field

[0001] The present invention belongs to the technical field of image recognition, and in particular relates to a method for recognizing and classifying kidney tumor MRI tomographic images based on image distribution mathematical characteristics and a support vector machine. Background Art

[0002] Image recognition refers to the use of computers to process, analyze, and understand image data to identify various patterns of targets and objects. To achieve more accurate image recognition and classification results, artificial intelligence and machine learning algorithms are needed to learn the patterns and features in images. Feature extraction primarily focuses on extracting the most representative image features from experimental data.

[0003] Existing medical image recognition methods primarily include those based on machine learning and deep learning. Machine learning is a data-driven approach that uses training data and experience to build mathematical models for data prediction and analysis. Deep learning, essentially a branch of machine learning, is a learning method based on neural networks. It can analyze medical images by learning and training on large amounts of medical image data.

[0004] Existing medical image recognition only extracts some basic imaging genomics features and some convolutional features that are not physically interpretable, and does not consider more complex and representative features. Summary of the Invention

[0005] The purpose of the present invention is to address the shortcomings of existing methods and propose a feature extraction and classification method for kidney tumor MRI tomography images to solve the problem of poor feature extraction and classification effect in medical image recognition.

[0006] A method for extracting and classifying features of MRI tomographic images of kidney tumors according to the present invention comprises the following steps:

[0007] 1) Using MRI scans of actual kidney tumor patients as experimental data and the different Ki67 expression results of the patients as classification criteria, a feature vector was established to classify the images, including the following steps:

[0008] 1.1 Obtain renal MRI T1-weighted images of patients with renal tumors as a training set;

[0009] 1.2 According to the proportion of positive cells, the data were divided into three categories according to the positive expression results of Ki67: weak positive, positive, and strong positive;

[0010] 1.3 Use grayscale equalization and other methods to preprocess the training set images;

[0011] 2) In feature extraction, considering the characteristics of renal tumor MRI tomography images, a Gaussian mixture model is used to fit the MRI tomography images to obtain mathematical features representing the image distribution. These mathematical features are then screened to obtain representative image features, including the following steps:

[0012] 2.1 Use K-means clustering algorithm to perform clustering and calculate the pixel cluster category. Its mathematical expression is:

[0013]

[0014] Among them: iClass is the j-th cluster center; μ j is the mean of all pixel vectors in the jth cluster;

[0015] 2.2 Using the clustering results, calculate the initial mean, initial covariance and initial prior probability by group;

[0016] 2.3 Using the parameters obtained in step 2.2, calculate the probability that each pixel belongs to a different category. The mathematical expression is:

[0017]

[0018] Where: X i is a random vector; Θ j is the set of distribution parameters of the jth distribution; μ j is the mean of all pixel vectors in the jth distribution; Σ j is the covariance matrix of the jth distribution; Dim is the dimension of the data;

[0019] 2.4 Using the calculated probability of each pixel belonging to a different category, we can classify each pixel into its own category. The mathematical expression is:

[0020]

[0021] 2.5 According to the prior probability α j Constraints it has: And the posterior probability ω ij satisfy The prior probability is calculated using the Lagrange multiplier method, and its mathematical expression is:

[0022]

[0023] Where: α j is the prior probability, that is, the probability that a pixel belongs to the jth distribution; ω ij is the posterior probability Φ(X i |Θ j), that is, based on the currently calculated parameters, the probability that the i-th pixel belongs to the j-th distribution is obtained; J is the logarithmic likelihood function, that is, the probability that all points belong to the current distribution, and its mathematical expression is: Λ is the parameter set of the Gaussian mixture model; η i is the Lagrange coefficient; N is the number of pixels; K is the number of distributions;

[0024] 2.6 According to the posterior probability ω ij , calculate the mean:

[0025]

[0026] Where: μ j is the mean of all pixel vectors in the jth distribution;

[0027] 2.7 Calculate the iterative covariance matrix:

[0028]

[0029] 2.8 Assume that the distribution of the image is a Gaussian mixture model and calculate the total probability density:

[0030]

[0031] Where: X is a random vector X i is a set of parameters of the Gaussian mixture model; p ΛY (i) is the probability density function of the i-th pixel; N is the number of pixels; K is the number of Gaussian distributions; α j is the prior probability; Φ(X i |Θ j ) is the probability density function of Gaussian distribution; Θ j are other parameters of the jth distribution;

[0032] 2.9 Repeat steps 2.3-2.8, and calculate the sum of the differences between the parameters of this iteration and the parameters of the previous iteration. If the difference is greater than 0.001, continue iterating. If the total probability density does not increase significantly and the sum of the differences is less than 0.001, the parameters of this iteration are the parameters of the Gaussian mixture model.

[0033] 2.10 The model parameters calculated in step 2.9 are used as mathematical eigenvectors, and feature screening is performed on this basis to retain the most representative eigenvectors. The mathematical expression of feature screening is:

[0034]

[0035] Among them: F iis the F-score of the i-th eigenvector; x is the training sample set; is the average value of the i-th feature in the j-th data set; is the average value of the i-th feature over the entire dataset; n j is the number of samples of class j; is the eigenvalue of the i-th feature of the k-th sample point of the j-th category; l is the number of training sample categories;

[0036] 3) Using a support vector machine to classify the features of different images and link kidney tumor MRI images with Ki67 expression to obtain more accurate image recognition and classification results, including the following steps:

[0037] 3.1 Using support vector machine, the Ki67 expression results were used as the classification criteria to train and test the feature vectors of different images. The decision boundary b ij The discriminant function is:

[0038]

[0039] Where: Experimental data is T={(x1,y1),(x2,y2),...,(x n ,y n )}; classification standard is y i ∈{c1,c2,...,c k}; w is the weight vector of each vector; φ represents the nonlinear mapping of the input feature space; b is the bias, that is, the displacement of the hyperplane relative to the origin; t is the index of the union of the data sets of i and j;

[0040] 3.2 Using SVM classification function to judge data x new The mathematical expression of the category is:

[0041]

[0042] Where: x new is the test sample set; α is the Lagrange coefficient; k is the kernel function; the radial basis Gaussian kernel function used in the low-dimensional feature space calculation is:

[0043]

[0044] Where: x i It is the i-th eigenvector of the input, which mainly measures the similarity between x1 and x2 in space.

[0045] The characteristics and beneficial effects of the present invention are that, compared with existing algorithms, this invention not only draws on the ideas of traditional image recognition methods but also draws on the image features inherent in MRI tomography images of patients with kidney tumors to achieve more accurate image recognition and classification results. By fitting the image data with a Gaussian mixture model, mathematical features representing the image distribution are obtained, enriching the types of MRI tomography image features used in existing machine learning methods, making the image features both representative and physically interpretable. In clinical medicine, this invention can assist physicians in determining the pathological stage and malignant tendency of patients' kidney tumors. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A flowchart of a feature extraction and classification method for renal tumor MRI tomography images;

[0047] Figure 2 The original image and clustered image of MRI tomography of kidney tumor;

[0048] Among them: (a) is the original image, (b) is the image after clustering;

[0049] Figure 3 A scatter diagram of all the mathematical characteristics of each model in the Gaussian mixture model;

[0050] Figure 4 The figure is a scatter diagram of the mathematical features of six different positions of the Gaussian mixture model;

[0051] Figure 5 This is the image recognition and classification results obtained using support vector machines;

[0052] Figure 3 、 Figure 4 and Figure 5 Middle: (a) represents the two-classification case, and (b) represents the three-classification case. DETAILED DESCRIPTION

[0053] A method for extracting and classifying features of MRI tomographic images of kidney tumors according to the present invention comprises the following steps:

[0054] 1) Using MRI scans of actual kidney tumor patients as experimental data and the different Ki67 expression results of the patients as classification criteria, a feature vector was established to classify the images, including the following steps:

[0055] 1.1 Obtain renal MRI T1-weighted images of patients with renal tumors as a training set;

[0056] 1.2 According to the proportion of positive cells, the data were divided into three categories according to the positive expression results of Ki67: weak positive, positive, and strong positive;

[0057] 1.3 Use grayscale equalization and other methods to preprocess the training set images;

[0058] 2) In feature extraction, considering the characteristics of renal tumor MRI tomography images, a Gaussian mixture model is used to fit the MRI tomography images to obtain mathematical features representing the image distribution. These mathematical features are then screened to obtain representative image features, including the following steps:

[0059] 2.1 Use K-means clustering algorithm to perform clustering and calculate the pixel cluster category. Its mathematical expression is:

[0060]

[0061] Among them: iClass is the j-th cluster center; μ j is the mean of all pixel vectors in the jth cluster;

[0062] 2.2 Using the clustering results, calculate the initial mean, initial covariance and initial prior probability by group;

[0063] 2.3 Using the parameters obtained in step 2.2, calculate the probability that each pixel belongs to a different category. The mathematical expression is:

[0064]

[0065] Where: X i is a random vector; Θ j is the set of distribution parameters of the jth distribution; μ j is the mean of all pixel vectors in the jth distribution; Σ j is the covariance matrix of the jth distribution; Dim is the dimension of the data;

[0066] 2.4 Using the calculated probability of each pixel belonging to a different category, we can classify each pixel into its own category. The mathematical expression is:

[0067]

[0068] 2.5 According to the prior probability α j Constraints it has: And the posterior probability ω ij satisfy The prior probability is calculated using the Lagrange multiplier method, and its mathematical expression is:

[0069]

[0070] Where: α j is the prior probability, that is, the probability that a pixel belongs to the jth distribution; ωij is the posterior probability Φ(X i |Θ j ), that is, based on the currently calculated parameters, the probability that the i-th pixel belongs to the j-th distribution is obtained; J is the logarithmic likelihood function, that is, the probability that all points belong to the current distribution, and its mathematical expression is: Λ is the parameter set of the Gaussian mixture model; η i is the Lagrange coefficient; N is the number of pixels; K is the number of distributions;

[0071] 2.6 According to the posterior probability ω ij , calculate the mean:

[0072]

[0073] Where: μ j is the mean of all pixel vectors in the jth distribution;

[0074] 2.7 Calculate the iterative covariance matrix:

[0075]

[0076] 2.8 Assume that the distribution of the image is a Gaussian mixture model and calculate the total probability density:

[0077]

[0078] Where: X is a random vector X i is a set of parameters of the Gaussian mixture model; p ΛY (i) is the probability density function of the i-th pixel; N is the number of pixels; K is the number of Gaussian distributions; α j is the prior probability; Φ(X i |Θ j ) is the probability density function of Gaussian distribution; Θ j are other parameters of the jth distribution;

[0079] 2.9 Repeat steps 2.3-2.8, and calculate the sum of the differences between the parameters of this iteration and the parameters of the previous iteration. If the difference is greater than 0.001, continue iterating. If the total probability density does not increase significantly and the sum of the differences is less than 0.001, the parameters of this iteration are the parameters of the Gaussian mixture model.

[0080] 2.10 The model parameters calculated in step 2.9 are used as mathematical eigenvectors, and feature screening is performed on this basis to retain the most representative eigenvectors. The mathematical expression of feature screening is:

[0081]

[0082] Among them: F i is the F-score of the i-th eigenvector; x is the training sample set; is the average value of the i-th feature in the j-th data set; is the average value of the i-th feature over the entire dataset; n j is the number of samples of class j; is the eigenvalue of the i-th feature of the k-th sample point of the j-th category; l is the number of training sample categories;

[0083] 3) Using a support vector machine to classify the features of different images and link kidney tumor MRI images with Ki67 expression to obtain more accurate image recognition and classification results, including the following steps:

[0084] 3.1 Using support vector machine, the Ki67 expression results were used as the classification criteria to train and test the feature vectors of different images. The decision boundary b ij The discriminant function is:

[0085]

[0086] Where: Experimental data is T={(x1,y1),(x2,y2),...,(x n ,y n )}; classification standard is y i ∈{c1,c2,...,c k}; w is the weight vector of each vector; φ represents the nonlinear mapping of the input feature space; b is the bias, that is, the displacement of the hyperplane relative to the origin; t is the index of the union of the data sets of i and j;

[0087] 3.2 Using SVM classification function to judge data x new The mathematical expression of the category is:

[0088]

[0089] Where: x new is the test sample set; α is the Lagrange coefficient; k is the kernel function; the radial basis Gaussian kernel function used in the low-dimensional feature space calculation is:

[0090]

[0091] Where: x i It is the i-th eigenvector of the input, which mainly measures the similarity between x1 and x2 in space.

[0092] The following specific tests verify the feasibility of the feature extraction and classification method for kidney tumor MRI tomography images of the present invention. The kidney tumor MRI tomography image recognition and classification results of the method of the present invention are compared with the image recognition method proposed by Carnati Emanuele, and the recognition and classification results of different categories of kidney tumor MRI tomography images of the present invention are demonstrated.

[0093] 1. Working conditions:

[0094] The experimental platform of the present invention adopts Intel(R) Core(TM) i5-8600K CPU@3.60GHz, 16.0GB memory, a PC running Windows 10, and the programming language is MATLAB.

[0095] 2. Experimental content and results analysis:

[0096] Table 1 compares the classification results of the mathematical features obtained by fitting the image data by the Gaussian mixture model in the present invention with the classification results of the features such as morphological features, grayscale histogram features, grayscale co-occurrence matrix features, and grayscale scale area matrix features extracted by the traditional method. The two-category accuracy is the ratio of the number of samples correctly classified in the test set to the total number of samples after the image data training set is divided into two categories; the three-category accuracy is the ratio of the number of samples correctly classified in the test set to the total number of samples after the image data training set is divided into three categories. This accuracy value can be used to evaluate the representativeness of the extracted features. It can be seen from Table 1 that the two-category accuracy of the method of the present invention is 25.7147% higher than that of the reference method, and the three-category accuracy of the method of the present invention is 19.4633% higher than that of the reference method.

[0097] like Figure 2 As shown, Figure 2 This is a comparison diagram of the original MRI tomography image of kidney tumor and the image after clustering. It can be seen that the MRI tomography image of kidney tumor has obvious characteristics in distribution characteristics.

[0098] like Figure 3 As shown, Figure 3 (a) is a scatter diagram of all the mathematical features of each model in the Gaussian mixture model in the case of binary classification. Figure 3 (b) is a scatter diagram of all the mathematical features of each model in the Gaussian mixture model under the three-classification condition. Figure 4 As shown, Figure 4 (a) is a scatter diagram of the mathematical features at six different positions of the Gaussian mixture model in the case of binary classification. Figure 4(b) is a scatter diagram of the mathematical features at six different positions of the Gaussian mixture model in the case of three classifications.

[0099] like Figure 5 As shown, Figure 5 (a) is the actual classification and predicted classification diagram of the test set in the case of binary classification. Figure 5 (b) is the actual classification and predicted classification diagram of the test set in the case of three classifications.

[0100] Table 1 Image recognition and classification results

[0101] Method of the present invention Reference Methods Binary classification accuracy 83.1615％ 57.4468％ Three-category accuracy 79.0378％ 59.5745％

[0102] The experimental results above demonstrate that the present invention uses a Gaussian mixture model to fit image data, extracting mathematical features representing image distribution. Using a support vector machine to classify the features of different images, the present invention links kidney tumor MRI images with Ki67 expression, resulting in more accurate image recognition and classification results. Furthermore, the present invention possesses physical interpretability and can assist clinicians in clinically assessing the pathological stage and malignant propensity of kidney tumors.

Claims

1. A method for feature extraction and classification of renal tumor MRI tomographic images, comprising the following steps: 1) Using MRI scans of actual kidney tumor patients as experimental data and the different Ki67 expression results of the patients as classification criteria, a feature vector was established to classify the images, including the following steps: 1.1 Obtain renal MRI T1-weighted images of patients with renal tumors as a training set; 1.2 According to the proportion of positive cells, the data were divided into three categories according to the positive expression results of Ki67: weak positive, positive, and strong positive; 1.3 Use grayscale equalization method to preprocess the training set images; 2) In feature extraction, considering the characteristics of renal tumor MRI tomography images, a Gaussian mixture model is used to fit the MRI tomography images to obtain mathematical features representing the image distribution. These mathematical features are then screened to obtain representative image features, including the following steps: 2.1 Use K-means clustering algorithm to perform clustering and calculate the pixel cluster category. Its mathematical expression is: in: iClass is the jth cluster center; μ j is the mean of all pixel vectors in the jth cluster; 2.2 Using the clustering results, calculate the initial mean, initial covariance and initial prior probability by group; 2.3 Using the parameters obtained in step 2.2, calculate the probability that each pixel belongs to a different category. The mathematical expression is: Where: X i is a random vector; Θ j is the set of distribution parameters of the jth distribution; μ j is the mean of all pixel vectors in the jth distribution; Σ j is the covariance matrix of the jth distribution; Dim is the dimension of the data; 2.4 Using the calculated probability of each pixel belonging to a different category, we can classify each pixel into its own category. The mathematical expression is: 2.5 According to the prior probability α j Constraints it has: And the posterior probability ω ij satisfy The prior probability is calculated using the Lagrange multiplier method, and its mathematical expression is: in: α j is the prior probability, that is, the probability that a pixel belongs to the jth distribution; ω ij is the posterior probability Φ(X i |Θ j ), that is, based on the currently calculated parameters, the probability that the i-th pixel belongs to the j-th distribution is obtained; J is the logarithmic likelihood function, that is, the probability that all points belong to the current distribution, and its mathematical expression is: Λ is the set of parameters of the Gaussian mixture model; η i is the Lagrange coefficient; N is the number of pixels; K is the number of distributions; 2.6 According to the posterior probability ω ij , calculate the mean: Where: μ j is the mean of all pixel vectors in the jth distribution; 2.7 Calculate the iterative covariance matrix: 2.8 Assume that the distribution of the image is a Gaussian mixture model and calculate the total probability density: in: X is a random vector X i is a set of parameters of the Gaussian mixture model; p ΛY (i) is the probability density function of the i-th pixel; N is the number of pixels; K is the number of Gaussian distributions; α j is the prior probability; Φ(X i |Θ j ) is the probability density function of Gaussian distribution; Θ j are other parameters of the jth distribution; 2.9 Repeat steps 2.3-2.8, and calculate the sum of the differences between the parameters of this iteration and the parameters of the previous iteration. If the difference is greater than 0.001, continue iterating. If the total probability density does not increase significantly and the sum of the differences is less than 0.001, the parameters of this iteration are the parameters of the Gaussian mixture model. 2.10 The model parameters calculated in step 2.9 are used as mathematical eigenvectors, and feature screening is performed on this basis to retain the most representative eigenvectors. The mathematical expression of feature screening is: Among them: F i is the F-score of the i-th eigenvector; x is the training sample set; is the average value of the i-th feature in the j-th data set; is the average value of the i-th feature over the entire dataset; n j is the number of samples of class j; is the eigenvalue of the i-th feature of the k-th sample point of the j-th category; l is the number of training sample categories; 3) Using a support vector machine to classify the features of different images and link kidney tumor MRI images with Ki67 expression to obtain more accurate image recognition and classification results, including the following steps: 3.1 Using support vector machine, the Ki67 expression results were used as the classification criteria to train and test the feature vectors of different images. The decision boundary b ij The discriminant function is: Where: Experimental data is T={(x1,y1),(x2,y2),...,(x n ,y n )}; classification standard is y i ∈{c1,c2,...,c k }; w is the weight vector of each vector; φ represents the nonlinear mapping of the input feature space; b is the bias, that is, the displacement of the hyperplane relative to the origin; t is the index of the union of the data sets of i and j; 3.2 Using SVM classification function to judge data x new The mathematical expression of the category is: Where: x new is the test sample set; α is the Lagrange coefficient; k is the kernel function; the radial basis Gaussian kernel function used in the low-dimensional feature space calculation is: Where: x i It is the i-th eigenvector of the input, which measures the similarity between x1 and x2 in space.

Images