A Medical Image Classification Method Based on Lie Group Kernel Learning

Abstract

This invention belongs to the technical field of image classification and discloses a medical image classification method based on Lie group kernel learning. The method includes acquiring low-level image features and representing them as a Lie group matrix; obtaining the model parameters of an SVM classifier and the pivot point for each class through a training image set; and selecting either an SVM or KNN classifier to classify the images based on the geodesic distance between the class pivot point and each image to be classified, calculated using the Lie group kernel function on the Lie group manifold. This invention outperforms traditional image classification methods in terms of classification accuracy. Compared to artificial neural network methods, it has less dependence on training data and processes, and exhibits stronger generalization and deployability.

Description

Technical Field

[0001] This invention relates to the technical field of image classification, and more particularly to a medical image classification method based on Lie group kernel learning. Background Technology

[0002] Thanks to the rapid development of digital medical imaging in modern medicine, the accessibility and accuracy of medical images are increasing, and medical image analysis is playing an increasingly important role in most computer-aided diagnoses, medical image retrieval, and medical data mining. Medical image classification is an indispensable step in medical image analysis, aiming to differentiate medical images based on certain criteria (such as clinical pathology or imaging modality). Medical images often contain a great deal of texture, making classification using simple observation or traditional classification methods unreliable, or even impossible. A specialized method capable of producing accurate classifications and robust to different medical imaging modalities is crucial for the rapid and accurate interpretation of medical images.

[0003] Medical image classification has been extensively studied using various methods. Research in the literature generally falls into one of two categories: traditional methods based on human-made features and deep learning-based methods using CNNs.

[0004] Artificial features refer to manually designed image attributes derived from the image itself using information presented within the image. Handcrafted feature representations use one or more low-level image features (such as color, texture, and shape) as descriptors to detect salient regions of an image. Typically, traditional methods combine handcrafted feature representations with standard classifiers such as SVM, Principal Component Analysis (PCA), and Linear Discriminant Analysis (LDA). Three limitations of these traditional methods include: (1) heavy reliance on prior empirical information; (2) neglect of relationships between different regions and features in the image; and (3) sensitivity to large rotations and drastic changes in illumination.

[0005] Deep learning-based methods provide a unified framework for feature extraction and classification. A hierarchical mechanism is employed to extract features, transforming the extracted features into more abstract and higher-level features. When applied to medical image classification and analysis, existing deep learning-based methods often have two main limitations: (1) intra-class variations and inter-class similarities are highly sensitive to the diversity of imaging methods and clinical pathologies; (2) they often ignore local region information in images because they always have too many model parameters to explain the learning of lower-level features, instead using only the output of the final connection layer of the network model as a global feature representation for classification prediction. Sometimes, a method may achieve optimal classification results on one medical image dataset but fail on another. Summary of the Invention

[0006] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0007] In view of the aforementioned existing problems, the present invention is proposed.

[0008] Therefore, the technical problem solved by the present invention is that the existing technology is unreliable for medical image classification.

[0009] To address the aforementioned technical problems, this invention provides the following technical solution: a medical image classification method based on Lie group kernel learning, comprising:

[0010] Obtain the low-level image features of the image and represent them as a Lie group matrix based on the image features;

[0011] The model parameters of the SVM classifier and the pivot for each class are obtained by training the image set;

[0012] Based on the calculation of the geodesic distance between the class pivot and each image to be classified on the Lie group manifold using the Lie group kernel function, either SVM or KNN classifier is selected to classify the image.

[0013] As a preferred embodiment of the medical image classification method based on Lie group kernel learning described in this invention, the low-level image features include colors described by RGB and YCR models, first-order and second-order gradients, and textures described by Gabor and LBP filters.

[0014] Calculate the low-level image features for each pixel (x, y) of an image;

[0015] The formula for calculating the Gabor feature is as follows:

[0016]

[0017] Where, ω x Let σ be the frequency of the sine wave, σ be the standard deviation of the Gaussian function in the x and y directions, and θ be the direction of the filter.

[0018] As a preferred embodiment of the medical image classification method based on Lie group kernel learning described in this invention, wherein: the image pixels after calculating the low-level features are represented as feature vectors, and the feature vectors are represented as follows:

[0019]

[0020] in, and It is the first-order gradient along the X and Y directions. and It is the second gradient along the X and Y directions. It is the sum of squares of the first-order gradients. It is the arctangent function of the first-order gradient ratio.

[0021] As a preferred embodiment of the medical image classification method based on Lie group kernel learning described in this invention, the covariance features of the image are obtained by statistically analyzing the combination of multiple relevant features of the image, wherein the covariance features include:

[0022] The covariance of a series of d-channel eigenvectors is a d×d matrix, where the diagonal terms represent the variance of each eigenvector and the off-diagonal terms represent the attribute correlation between each pair of vectors.

[0023] For an image with N pixels, the covariance matrix C is expressed as:

[0024]

[0025] Among them, f k Let μ be a d-dimensional eigenvector, and μ be the average vector of all eigenvectors, expressed as:

[0026]

[0027] As a preferred embodiment of the medical image classification method based on Lie group kernel learning described in this invention, wherein: the covariance matrix C is subjected to Cholesky decomposition to obtain the Lie group SOG matrix features, and the Lie group SOG matrix features are expressed as:

[0028]

[0029] Where R is the Cholesky decomposition of C, and C = RR T , is an upper triangular matrix with a special structure.

[0030] As a preferred embodiment of the medical image classification method based on Lie group kernel learning described in this invention, the support matrix of the SVM model is obtained by training the image set, and the distance between the SOG matrices Mi and Mj of two images on the manifold is expressed as:

[0031] D LG (M i M j )=||log2(M j )-log2(M i )|| 2

[0032] The similarity between two images is measured using the Lie group RBF kernel function, and is expressed as follows:

[0033]

[0034] Where γ is the scaling factor.

[0035] As a preferred embodiment of the medical image classification method based on Lie group kernel learning described in this invention, in an SVM model, a data sample can be distinguished by a hyperplane, which must satisfy the following condition:

[0036]

[0037] The conditions are transformed into a Lagrange optimization equation using Lagrange operators, as follows:

[0038]

[0039] Where, α i For Lagrange operators, Let y be the SOG matrix of an image sample in the training set, where n is the size of the training set, and y is the size of the training set. i This is the category label, with a value of 1 or -1;

[0040] The Lagrange optimization equation, differentiated with respect to ω and b respectively and set the result to 0, is expressed as follows:

[0041]

[0042]

[0043] As a preferred embodiment of the medical image classification method based on Lie group kernel learning described in this invention, the Lagrange optimization equation is rewritten as follows: By utilizing the matrix similarity relation expressed by the kernel function and the result of differentiation, the Lagrange optimization equation is rewritten as:

[0044]

[0045]

[0046] Where C is a penalty coefficient;

[0047] The optimal solution of the rewritten Lagrange optimization equation is expressed as:

[0048]

[0049] The two parameters of the SVM model discriminant equation are respectively expressed as:

[0050]

[0051]

[0052] The support matrix for each category is represented as follows:

[0053]

[0054] As a preferred embodiment of the medical image classification method based on Lie group kernel learning described in this invention, wherein: the distance from each image to be classified to the positive class pivot M is measured respectively. + and negative class pivot M - The geodesic distance is calculated, and then the difference between the two distances is calculated, which is expressed as:

[0055]

[0056]

[0057] When dif(M) is greater than or equal to the threshold, the image to be classified is far from the hyperplane, and the SVM classifier is selected.

[0058] When dif(M) is less than the threshold, the image to be classified is close to the hyperplane, and the KNN classifier is selected.

[0059] As a preferred embodiment of the medical image classification method based on Lie group kernel learning described in this invention, when an SVM classifier is selected for image classification, the matrix is ​​M. x The category of the image to be tested is determined by a determination function, which is expressed as:

[0060]

[0061] When the KNN classifier is selected to classify an image, the test image M is calculated. x To each category of pivot point M g The distances are calculated, and then all distances are sorted in ascending order and stored in an array. Finally, the categories corresponding to the first K distances in the array are selected to classify the image to be tested. The distances are represented as:

[0062]

[0063] The beneficial effects of this invention are as follows: This invention provides a medical image classification method based on Lie group kernel learning. By employing Lie group features of SOG to encode the color, gradient, shape features, and association information of the image, it forms a compact medical image representation, effectively addressing the current method's deficiency of "ignoring the relationships between different regions and features in the image." Mapping the image onto a Lie group manifold and using the Lie group kernel function to measure the similarity between different images effectively addresses the current method's sensitivity to "large rotations and drastic lighting changes." Combining the SVM classifier with the KNN classifier improves the classifier's accuracy and solves the problems of imbalanced classification near the SVM hyperplane and the high misclassification rate of the KNN classifier. This invention outperforms traditional image classification methods in terms of classification accuracy. Compared to artificial neural network methods, it has less dependence on training data and processes, and exhibits stronger generalization and deployability. Attached Figure Description

[0064] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0065] Figure 1 This is an overall flowchart of a medical image classification method based on Lie group kernel learning according to an embodiment of the present invention;

[0066] Figure 2 This is a schematic diagram of a Lie group kernel-based medical image classification method according to an embodiment of the present invention.

[0067] Figure 3 This is a schematic diagram showing the distance between data samples and the positive and negative class pivots of the medical image classification method based on Lie group kernel learning according to an embodiment of the present invention;

[0068] Figure 4 The diagram illustrates the performance of the classifier using ROC curves for a medical image classification method based on Lie group kernel learning, as described in one embodiment of the present invention.

[0069] Figure 5 This is a schematic diagram of the confusion matrix and ROC curve of a medical image classification method based on Lie group kernel learning according to an embodiment of the present invention. Detailed Implementation

[0070] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0071] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0072] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0073] This invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of this invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not adhering to the usual scale. Furthermore, the schematic diagrams are merely examples and should not be construed as limiting the scope of protection of this invention. In actual fabrication, the three-dimensional spatial dimensions of length, width, and depth should be included.

[0074] Furthermore, in the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used solely for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In addition, the terms "first," "second," or "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0075] Unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" in this invention should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; similarly, they can refer to mechanical connections, electrical connections, or direct connections, or indirect connections through an intermediate medium, or internal connections between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0076] Example 1

[0077] Reference Figure 1—3, is the first embodiment of the present invention, which provides a medical image classification method based on Lie group kernel learning, characterized in that it includes:

[0078] S1: Obtain the low-level image features of the image and represent them as a Lie group matrix based on the image features;

[0079] Furthermore, low-level image features include colors described by RGB and YCR models, first-order and second-order gradients, and textures described by Gabor and LBP filters.

[0080] Calculate the low-level image features for each pixel (x, y) of an image;

[0081] The formula for calculating the Gabor feature is as follows:

[0082]

[0083] Where, ω x Let σ be the frequency of the sine wave, σ be the standard deviation of the Gaussian function in the x and y directions, and θ be the direction of the filter.

[0084] It should be noted that the LBP feature records the difference between neighboring pixels and the center pixel. Therefore, when changes in illumination cause the grayscale values ​​of pixels within the window to increase or decrease synchronously, the LBP value does not change significantly. Thus, LBP is not sensitive to changes in illumination.

[0085] Furthermore, the image pixels after calculating low-level features are represented as feature vectors, as shown in the following feature vector representation:

[0086]

[0087] in, and It is the first-order gradient along the X and Y directions. and It is the second gradient along the X and Y directions. It is the sum of squares of the first-order gradients. It is the arctangent function of the first-order gradient ratio.

[0088] Furthermore, for grayscale images, the six components of the color model in the above formula are replaced with the intensity values ​​of the pixels.

[0089] Furthermore, the covariance features of the image are obtained by statistically analyzing the combination of multiple relevant features of the image, and the covariance features include:

[0090] The covariance of a series of d-channel eigenvectors is a d×d matrix, where the diagonal terms represent the variance of each eigenvector and the off-diagonal terms represent the attribute correlation between each pair of vectors.

[0091] For an image with N pixels, the covariance matrix C is expressed as:

[0092]

[0093] Among them, f k Let μ be a d-dimensional eigenvector, and μ be the average vector of all eigenvectors, expressed as:

[0094]

[0095] Furthermore, Cholesky decomposition is performed on the covariance matrix C to obtain the features of the Lie group SOG matrix, which are expressed as follows:

[0096]

[0097] Where R is the Cholesky decomposition of C, and C = RR T , is an upper triangular matrix with a special structure.

[0098] It should be noted that, in order to take into account the relationship between different regions and features when representing an image, this invention constructs a Lie group feature called SOG, which is derived by calculating the covariance of the image feature vector.

[0099] S2: Obtain the model parameters of the SVM classifier and the pivot for each class by training the image set;

[0100] Furthermore, the support matrix of the SVM model is obtained by training the image set. The distance between the SOG matrices Mi and Mj of two images on the manifold is expressed as:

[0101] D LG (M i M j )=||log2(M j )-log2(M i )|| 2

[0102] The similarity between two images is measured using the Lie group RBF kernel function, and is expressed as follows:

[0103]

[0104] Where γ is the scaling factor.

[0105] It should be noted that since the images are all represented using SOG matrices, they are all on Lie group manifolds.

[0106] Furthermore, in an SVM model, a data sample can be distinguished by a hyperplane if the following condition is met:

[0107]

[0108] The conditions are transformed into a Lagrange optimization equation using Lagrange operators, as follows:

[0109]

[0110] Where, α i For Lagrange operators, Let y be the SOG matrix of an image sample in the training set, where n is the size of the training set, and y is the size of the training set. i This is the category label, with a value of 1 or -1;

[0111] The Lagrange optimization equation, differentiated with respect to ω and b respectively and set the result to 0, is expressed as follows:

[0112]

[0113]

[0114] Furthermore, by utilizing the matrix similarity relation expressed by the kernel function and the result of differentiation, the Lagrange optimization equation is rewritten as:

[0115]

[0116]

[0117] Where C is a penalty coefficient;

[0118] The optimal solution of the rewritten Lagrange optimization equation is expressed as:

[0119]

[0120] The two parameters of the SVM model discriminant equation are respectively expressed as:

[0121]

[0122]

[0123] The support matrix for each category is represented as follows:

[0124]

[0125] S3: Based on the geodesic distance between the class pivot and each image to be classified, calculated using the Lie group kernel function on the Lie group manifold, select either SVM or KNN classifier to classify the image.

[0126] Furthermore, each image to be classified is measured to the positive class pivot M.+ and negative class pivot M - The geodesic distance is calculated, and then the difference between the two distances is calculated, which is expressed as:

[0127]

[0128]

[0129] When dif(M) is greater than or equal to the threshold, the image to be classified is far from the hyperplane, and the SVM classifier is selected.

[0130] When dif(M) is less than the threshold, the image to be classified is close to the hyperplane, and the KNN classifier is selected.

[0131] It should be noted that when test data samples are concentrated near the optimal hyperplane, the SVM classifier is more likely to give incorrect predictions. To address this issue, the proposed method combines the SVM classifier with the KNN classifier to better utilize the information provided by the support matrix near the hyperplane. More specifically, the algorithm treats the SVM classifier as a 1NN classifier, selecting only one representative support point for each class on the manifold.

[0132] Furthermore, when choosing an SVM classifier to classify images, the matrix is ​​M. x The category of the image to be tested is determined by a determination function, which is expressed as:

[0133]

[0134] When the KNN classifier is selected to classify an image, the test image M is calculated. x To each category of pivot point M g The distances are calculated, and then all distances are sorted in ascending order and stored in an array. Finally, the categories corresponding to the first K distances in the array are selected to classify the image to be tested. The distances are represented as:

[0135]

[0136] Example 2

[0137] Reference Figure 4 —5 is an embodiment of the present invention, which provides a medical image classification method based on Lie group kernel learning. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through specific implementation methods and implementation effects.

[0138] The specific implementation of this embodiment is as follows: The method proposed in this patent has been successfully applied to the classification of three sets of medical images: chest X-ray images, retinal OCT images, and ocular ROP images, and the results have been compared with clinical diagnosis results. The experimental results show that the method proposed in this patent is superior to traditional image classification methods in terms of classification accuracy. Compared with artificial neural network methods, it has less dependence on training data and processes, and has stronger generalization and deployability.

[0139] (1) Chest X-ray image

[0140] This experiment first classified patients with pneumonia on chest X-rays into those with normal chest X-rays, and then further differentiated between viral and bacterial pneumonia. Table 1 lists the binary classification (normal vs. pneumonia) comparison results of all 10 methods. As shown in Table 1, for different indicators, the number of methods achieving above 90% were: accuracy 5, precision 2, sensitivity 5, specificity 6, F1 score 5, and AUC 9. Overall, the classification performance of the 4 methods in group 2 was comparable to that of the 5 methods in group 1, but the method presented in this study achieved the highest accuracy (96.3%), precision (92.7%), sensitivity (97.9%), specificity (95.4%), F1 score (96.6%), and AUC (99.4%) among the 10 methods.

[0141] Table 1. Binary classification results of chest X-ray images using different methods.

[0142] Group Methods Accuracy Precision Sensitivity Specificity F1 score AUC 1 SIFT+PCA 0.811 0.730 0.786 0.826 0.806 0.922 1 LBP+SVM 0.732 0.642 0.684 0.762 0.721 0.863 1 HOG+SVM 0.911 0.859 0.910 0.910 0.910 0.943 1 Gabor+SVM 0.902 0.847 0.902 0.903 0.902 0.962 1 Integrated+SVM 0.938 0.905 0.932 0.942 0.936 0.991 2 SOG+Lie-Mean 0.779 0.748 0.748 0.841 0.792 0.974 2 SOG+Lie-LDA 0.837 0.754 0.838 0.836 0.837 0.955 2 SOG+Lie-KDA 0.899 0.849 0.889 0.905 0.897 0.975 2 SOG+Lie-SVM 0.928 0.898 0.923 0.931 0.929 0.983 3 This invention 0.963 0.927 0.979 0.954 0.966 0.994

[0143] exist Figure 4 In this example, ROC curves are used to illustrate the performance of the classifier. Figure 4 As shown in Figure a, the ROC curve of the proposed method (black) is located at the outermost layer of the graph, achieving the highest AUC value (99.4%) among all methods, while the LBP+SVM method is located at the innermost layer (AUC: 87.3%). The area under the ROC curve distinguishing bacterial pneumonia from viral pneumonia and normal pneumonia is 92.5% (…). Figure 4 b) and 97.6% Figure 4 c) This indicates that the method can effectively detect pneumonia, thereby differentiating between viral and bacterial pneumonia, facilitating rapid referral of children requiring emergency intervention. From Figure 4 As shown in b and c, the AUC for normal vs. bacterial pneumonia is smaller than that for normal vs. viral pneumonia. This can be explained by the fact that bacterial pneumonia generally presents with milder symptoms compared to viral pneumonia. For comparison, the areas under the two ROC curves are reported to be 94% and 96.8%, respectively.

[0144] In summary, Table 1 and Figure 4The results demonstrate that the proposed method can effectively classify chest X-rays in pediatric patients, first identifying pneumonia and normal images to avoid misdiagnosis, and then distinguishing between viral and bacterial pneumonia, thereby helping children requiring urgent intervention or immediate antibiotic treatment. The performance of the proposed method is more attributable to the SVM-KNN classifier than to the SOG feature description. Furthermore, the proposed method exhibits similar overall performance quality to the deep learning-based method used in this experiment.

[0145] (2) Retinal OCT images

[0146] The classification results among CNV, DME, Drusen, and Normal are summarized in Table 2. Although Gabor+SVM and Integrated+SVM outperform SOG+Lie-Mean and SOG+Lie-LDA on some metrics, the five methods in Group 2 generally outperform the five conventional methods in classification performance. Furthermore, our proposed method outperforms the other methods on all six metrics: accuracy 95.4%, sensitivity 96.4%, specificity 95.1%, F1 score 95.7%, and Kappa score 94.2%. The only metric below 90% is accuracy, with a score of 86.7%.

[0147] Table 2 Classification results of retinal OCT images by different methods

[0148]

[0149]

[0150] To distinguish between the correct and incorrect classifications of CNV, DME, Drusen, and Normal, Figure 5 The correlation between predicted and true labels is described as a confusion matrix. The matrix shows that the proposed method can identify each category with high accuracy: 241 / 250 (96.4%) for Normal, 235 / 250 (94.0%) for Drusen, 240 / 250 (96.0%) for CNV, and 238 / 250 (95.2%) for DME. In contrast, deep learning-based methods generate confusion matrices on the same dataset: 246 / 250 (98.4%) for Normal, 235 / 250 (94.4%) for Drusen, 242 / 250 (96.8%) for CNV, and 237 / 250 (94.8%) for DME, slightly outperforming the proposed method in three out of four predictions.

[0151] In addition, three binary classifiers were implemented to distinguish Drusen, CNV, and DME images from normal images to observe the degradation of model performance. In the generated ROC curves: the first classifier (CNV vs. normal) achieved an AUC of 99.8%. Figure 5 b) The second classifier (DME vs. normal) achieved an AUC of 99.5%. Figure 5 c) The third classifier distinguished (Drusenvs were normal) and achieved an AUC of 99.2%. Figure 5 d). The right-angled shape of the ROC curve indicates that the classifier has near-perfect classification performance. The AUCs of the ROC curves are 100%, 99.87%, and 99.96%, respectively, all of which are higher than the method presented in this paper.

[0152] In this experiment, the Lie group kernel classifier was trained using a relatively small number of images (8,000 people, totaling 108,312), yet it still maintained high performance across all metrics, achieving satisfactory results and demonstrating the power of the Lie group kernel function. The classification results show that the SOG feature descriptor represents the key content of retinal OCT images more effectively than handcrafted feature descriptors, and the proposed SVM-KNN classifier further improves classification performance.

[0153] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A medical image classification method based on Lie group kernel learning, characterized in that, include: Obtain the low-level image features of the image and represent them as a Lie group matrix based on the image features; The model parameters of the SVM classifier and the pivot for each class are obtained by training the image set; Based on the calculation of the geodesic distance between the class pivot and each image to be classified on the Lie group manifold using the Lie group kernel function, either SVM or KNN classifier is selected to classify the image; The covariance features of an image are obtained by statistically analyzing a combination of multiple relevant features. These covariance features include: The covariance of a series of d-channel eigenvectors is a d×d matrix, where the diagonal terms represent the variance of each eigenvector and the off-diagonal terms represent the attribute correlation between each pair of vectors. For an image with N pixels, the covariance matrix C is expressed as: ； in, Let a be a d-dimensional eigenvector. The average vector of all eigenvectors is expressed as: ； Performing Cholesky decomposition on the covariance matrix C yields the SOG matrix features of the Lie group, which are expressed as follows: ； Where R is the Cholesky decomposition of C, , is an upper triangular matrix with a special structure; The support matrix of the SVM model is obtained by training the image set. The distance between the SOG matrices Mi and Mj of two images on the manifold is expressed as: ； The similarity between two images is measured using the Lie group RBF kernel function, and is expressed as follows: ； in, This is the scaling factor; Measure each image to be classified to the positive class pivot M. + and negative class pivot M - The geodesic distance is calculated, and then the difference between the two distances is calculated, which is expressed as: ； ； when When the threshold is greater than or equal to the threshold, the image to be classified is far from the hyperplane, and the SVM classifier is selected; when When the value is less than the threshold, the image to be classified is close to the hyperplane, and the KNN classifier is selected.

2. The medical image classification method based on Lie group kernel learning as described in claim 1, characterized in that: The low-level image features include colors described by RGB and YCR models, first-order and second-order gradients, and textures described by Gabor and LBP filters; Calculate the low-level image features for each pixel (x, y) of an image; The formula for calculating the Gabor feature is as follows: ； in, The frequency of the sine wave is... Let x be the standard deviation of the Gaussian function in the x and y directions. This indicates the direction of the filter.

3. The medical image classification method based on Lie group kernel learning as described in claim 2, characterized in that: The image pixels after calculating the low-level features are represented as feature vectors, as shown in the following feature vector representation: ； in, and It is the first-order gradient along the X and Y directions. It is the second gradient along the X and Y directions. It is the sum of squares of the first-order gradients. It is the arctangent function of the first-order gradient ratio.

4. The medical image classification method based on Lie group kernel learning as described in claim 3, characterized in that: In an SVM model, a data sample can be distinguished by a hyperplane if the following condition is met: ； The conditions are transformed into a Lagrange optimization equation using Lagrange operators, as follows: ； in, For Lagrange operators, Let n be the SOG matrix of an image sample in the training set, and n be the size of the training set. This is the category label, with a value of 1 or -1; The Lagrange optimization equation, by differentiating with respect to and and setting the result to 0, can be expressed as: 。 5. The medical image classification method based on Lie group kernel learning as described in claim 4, characterized in that: By utilizing the matrix similarity relation expressed by the kernel function and the result of differentiation, the Lagrange optimization equation is rewritten as: ； Where C is a penalty coefficient; The optimal solution of the rewritten Lagrange optimization equation is expressed as: ； The two parameters of the SVM model discriminant equation are respectively expressed as: ； The support matrix for each category is represented as follows: .

6. The medical image classification method based on Lie group kernel learning as described in claim 5, characterized in that, include: When an SVM classifier is chosen to classify images, the matrix is ​​M. x The category of the image to be tested is determined by a determination function, which is expressed as: ； When the KNN classifier is selected to classify an image, the test image M is calculated. x To each category of pivot point M g The distances are calculated, and then all distances are sorted in ascending order and stored in an array. Finally, the categories corresponding to the first K distances in the array are selected to classify the image to be tested. The distances are represented as: 。

Images