Medical focus identification method based on deep low-rank multi-label classification
By constructing a deep low-rank multi-label classification model based on cluster structure injection, the characterization ability of the lesion recognition model is enhanced by using near-neighbor samples and low-rank representation, the problem of difficult to capture the lesion connection with surrounding tissues is solved, the accuracy and robustness of medical lesion recognition are improved, and the multi-label classification task is adapted to the multi-label classification task.
Patent Information
- Application Number
- CN202510327600.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-19
- Publication Date
- 2025-07-04
AI Technical Summary
Existing medical lesion recognition technology is difficult to effectively capture the potential connection between lesion and surrounding tissues, and the recognition accuracy is not high when dealing with multi-label classification tasks, and the model generalization ability and robustness are insufficient in the face of complex and changeable medical image data.
A deep low-rank multi-label classification model based on cluster structure injection is constructed, and the data representation ability of the classifier is enhanced through nearest neighbor samples and low-rank representation. Combined with self-expression subspace clustering network and low-rank representation, the connection between traditional classifiers and low-rank representation is established, and the objective function is optimized to improve the classification performance and robustness of the model.
By fully digging out the potential characterization information between samples, the classification performance and robustness of lesion recognition are enhanced, the recognition accuracy of multiple lesions is improved, and the multi-label classification task is adapted to the multi-label classification task of complex medical images.
Smart Images
Figure CN120259749A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to image classification technology and the medical field, and specifically to a method for identifying medical lesions based on deep low-rank multi-label classification. Background Art
[0002] Medical lesion identification is a key link in medical image analysis and is crucial for the early detection and treatment of diseases. However, this field faces multiple challenges. The background of medical images is complex, lesions are often masked by surrounding tissue structures, and they exhibit diverse manifestations due to disease stages and individual differences. In addition, the image quality is affected by imaging equipment, scanning parameters, and patient status, increasing the difficulty of identification. Doctors need to have profound professional knowledge to accurately interpret images, but even with rich experience, they may encounter difficulties in interpretation due to the complexity of the images and time urgency. Therefore, the development of intelligent, efficient, and accurate lesion identification algorithms has become an urgent need. These algorithms need to overcome the complexity of images, improve the identification accuracy, and at the same time speed up the processing speed to adapt to emergency and dynamic monitoring scenarios. The progress of medical lesion identification will greatly promote the scientific nature of clinical decision-making and the effectiveness of patient treatment.
[0003] However, the existing medical lesion identification technologies still have some deficiencies. On the one hand, most methods only focus on the feature extraction of the lesions themselves and ignore the potential connections between the lesions and the surrounding tissues, which limits the further improvement of the identification accuracy. Information such as blood vessels and tissue morphology around the lesions often has important reference value for the nature and classification of the lesions, but traditional methods are often difficult to effectively capture this contextual information. On the other hand, in the face of complex and variable medical image data, the existing identification models perform poorly in dealing with multi-label classification tasks and are difficult to accurately identify multiple different types of lesions simultaneously. In addition, due to the high dimensionality and noise interference of medical image data, traditional feature extraction and classification methods are easily troubled by overfitting and high computational complexity, affecting the generalization ability and robustness of the model. Summary of the Invention
[0004] In view of the deficiencies in the prior art, the present invention provides a method for identifying medical lesions based on deep low-rank multi-label classification.
[0005] The present invention achieves the above technical objectives through the following technical means.
[0006] A method for identifying medical lesions based on deep low-rank multi-label classification, comprising:
[0007] Collect medical lesion image data and form a sample data set, divide the data set into a training set and a test set; construct a deep low-rank multi-label classification model based on clustering structure injection, and the deep low-rank multi-label classification model enhances the data representation ability of the original samples in the classifier through two parts: neighbor samples and low-rank representations; use the training set to train the deep low-rank multi-label classification model; use the trained deep low-rank multi-label classification model for medical lesion recognition.
[0008] Furthermore, each image in the data set is a sample, represented by a feature vector.
[0009] Furthermore, the division ratio of the training set to the test set is 6:4.
[0010] Furthermore, the construction process of the deep low-rank multi-label classification model is as follows:
[0011] First, use a linear regression model as the basic model of the multi-label classification model; second, construct neighbor samples to enhance the original samples through a self-expressive subspace clustering network, and at the same time add the self-expressive loss in the self-expressive subspace clustering network to the linear regression model to obtain an objective function; then, introduce low-rank representations to learn the low-rank space in the self-expressive subspace clustering network, and establish a constraint to build the connection between the traditional classifier and the low-rank representation, thereby updating the objective function; finally, perform regularization constraints on the updated objective function to construct the final deep low-rank multi-label classification model.
[0012] Furthermore, the self-expressive subspace clustering network is obtained by combining the self-expressive property and the subspace clustering network.
[0013] Furthermore, the formula of the objective function is:
[0014]
[0015] where is the data set reconstructed using neighbor samples, w is the parameter to be learned, Y is the label, ||·|| F is the Frobenius norm, λ1 is a hyperparameter, X is the sample data set, θ s is the parameter of the network self-expressive layer, I represents the identity matrix, and γ is the weight coefficient.
[0016] Furthermore, the formula of the updated objective function is as follows:
[0017]
[0018] where is the data set reconstructed using neighbor samples, w is the parameter to be learned, Y is the label, ||·||F is the F-norm, λ1 is a hyperparameter, p is the low-rank representation matrix, X is the sample data set, and θ s is the parameter of the network self-expression layer, I represents the identity matrix, γ is the weight coefficient, and q is the weight matrix for classification in the low-dimensional subspace.
[0019] Furthermore, the formula for the final deep low-rank multi-label classification model is:
[0020]
[0021] where is the data set reconstructed using the nearest neighbor samples, w is the parameter to be learned, ||·|| F is the F-norm, Y is the label, λ1, λ2, λ3, and λ4 are all manually set hyperparameters, p is the low-rank representation matrix, q is the weight matrix for classification in the low-dimensional subspace, X is the sample data set, and θ s is the parameter of the network self-expression layer, γ is the weight coefficient, and I is the identity matrix.
[0022] Furthermore, the training parameters of the multi-label classification model include the parameter θ of the network self-expression layer s , the parameter w to be learned, the low-rank representation matrix p, and the weight matrix q for classification in the low-dimensional subspace.
[0023] Furthermore, the evaluation metrics of the multi-label classification model include Example-F1 and Hamming loss, and the formulas are as follows:
[0024]
[0025]
[0026] where p i is the precision of the sample, r i is the recall rate of the sample, f(x i ) is the predicted label set of the multi-label classification of the real column x i , Y i is the label set, and Δ is the symmetric difference set operation.
[0027] Advantages of the present invention:
[0028] (1) The subspace clustering network of the present invention replaces the traditional feature design to mine the potential connections between samples, can fully mine the potential representation information between samples, construct semantically rich nearest neighbor samples to enhance the representation information of the original samples, and improve the classification performance.
[0029] (2) The present invention establishes a connection between a traditional classifier and a low-rank representation, enabling the classifier to not only enhance the original representation through the constructed neighboring samples but also further improve the classification robustness with the aid of the low-rank representation in the network. Brief Description of the Drawings
[0030] Figure 1 is a flowchart of the medical lesion recognition method for deep low-rank multi-label classification according to this embodiment.
[0031] Figure 2 is an architecture diagram of the deep low-rank multi-label classification model based on clustering structure injection according to this embodiment. Detailed Embodiments
[0032] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but the protection scope of the present invention is not limited thereto.
[0033] As Figure 1 shown, a medical lesion recognition method for deep low-rank multi-label classification includes the following steps:
[0034] Step 1: Prepare medical lesion image data. Each image is a sample, represented by a feature vector, to form a sample data set X ∈ R d*n , where n represents the number of samples and d represents the vector dimension. And perform professional annotation on the data in the data set, convert the image lesion information into a label Y. If the lesion exists, the label is 1; if the lesion does not exist, the label is 0, that is, Y ∈ {0, 1} n*m , where m represents the number of labels. The data set is divided into a training set and a test set according to a ratio of 6:4.
[0035] Step 2: Construct an architecture of a deep low-rank multi-label classification model based on clustering structure injection.
[0036] The deep low-rank multi-label classification model of this embodiment enhances the data representation ability of the original samples in the classifier through two parts: neighboring samples and low-rank representations, and then performs multi-label classification. As Figure 2 shown, the specific architecture of the multi-label classification model is as follows:
[0037] S1. The main task of multi-label classification is to learn a classifier that maps the sample feature space to the label space. The classifier can be obtained through the following linear regression model:
[0038]
[0039] where w is the parameter to be learned, and ||·|| F represents the Frobenius norm.
[0040] S2. According to the self-expression property, for any sample x in the sample data set X i, can be linearly represented by other samples in the same subspace, formalized as:
[0041] x i = X -i c i (2)
[0042] Among them, X -i represents the nearest neighbor dataset of sample x i .
[0043] However, in actual situations, samples are vulnerable to noise and cannot be directly represented by other samples in the same subspace. Therefore, a regularization constraint is added to the self-expression property to obtain the self-expression objective function:
[0044]
[0045] By solving the self-expression objective function, the linear combination coefficient c i is obtained, and then the nearest neighbor samples X i of sample x i are obtained using the linear combination coefficient c -i c i . The linear combination coefficients of all samples are constructed into a coefficient matrix C, and the diagonal elements of the coefficient matrix C are zero.
[0046] In addition, because the learned nearest neighbor samples X -i c i and the original samples x i have similar characteristic semantics, the original samples can be enhanced by the nearest neighbor samples to obtain more classification features, and the new samples can be expressed as:
[0047]
[0048] Among them, γ represents the weight coefficient.
[0049] Using the above sample data reconstruction method, the original sample x is replaced with the new sample i , and a new sample dataset is obtained. To improve the representation ability of the coefficient matrix C for samples, a subspace clustering network is further introduced to construct a self-expression subspace clustering network, which uses formula (3) as the loss function for recursive optimization. Therefore, the parameter matrix θ s of the network self-expression layer can be directly used to replace the coefficient matrix C, so as to obtain the nearest neighbor samples, and the self-expression loss in the self-expression subspace clustering network is added to the linear regression model to obtain the following objective function:
[0050]
[0051] Among them, is the dataset reconstructed using neighboring samples, I represents the identity matrix, and λ1 is a hyperparameter set manually.
[0052] S3. In high-dimensional data processing methods, low-rank representation is generally considered an effective method for dealing with noise, outliers, and redundant data in the data. Therefore, in this embodiment, a low-rank space in the self-expressive subspace clustering network is learned by introducing low-rank representation, so as to obtain the self-expression loss of the original samples and weighted samples in the low-rank space. Its expression form is as follows:
[0053]
[0054] Among them, p is the low-rank representation matrix.
[0055] Next, by establishing the constraint w = pq to construct the connection between the traditional classifier and the low-rank representation, q represents the weight matrix for classification in the low-dimensional subspace. This constraint can compress and encode the sample feature space. Further assume that the low-rank representation matrix p is uncorrelated, and impose the constraint p T p = I. Finally, the following objective function can be obtained:
[0056]
[0057] Perform regularization constraints on formula (7) to construct the final deep low-rank multi-label classification model. The model formula is as follows:
[0058]
[0059] Among them, λ2, λ3, and λ4 are all hyperparameters set manually to balance the model.
[0060] Step 3: Combine the attached Figure 1 , and use the training set in Step 1 to train the deep low-rank multi-label classification model constructed in Step 2 to obtain the optimal parameters. The training process is as follows:
[0061] Step 3.1: Solve the variable θ by optimizing formula (8) as the loss function s , and the loss function formula is as follows:
[0062]
[0063] Obtain the current optimal result: θ s .
[0064] Step 3.2: Iteratively solve the model parameters w, p, q:
[0065] By only retaining the terms containing w in formula (8) and combining the Lagrange multiplier method to solve the variable w:
[0066]
[0067] Obtained: w = (XX T + 2γXθ s X T + γ 2 Xθ s T θ s X T + λI n ) -1 (γXθ s Y + XY - Y1 + μ1pq), where I n is an n*n identity matrix, and tr() is the sum of the traces of the matrix; Y1 and μ1 are both Lagrange multipliers.
[0068] By only retaining the terms related to p in formula (8) and combining with the Lagrange multiplier method to solve the variable p:
[0069]
[0070] Let
[0071] Let
[0072] Obtained: p = U[I; 0]V T , where α is a hyperparameter; U and V are both n*n matrices, which are the singular value decomposition of matrix G.
[0073] By only retaining the terms related to q in formula (8) and combining with the Lagrange multiplier method to solve the variable q:
[0074]
[0075] Obtained:
[0076] After obtaining the preliminary results of θ s , w, p, and q through steps 3.1 and 3.2, then through backpropagation and iterative update until the loss function converges, the training is stopped, and finally the optimal parameters are obtained.
[0077] Step 4: Use the trained deep low-rank multi-label classification model for medical lesion recognition.
[0078] To verify the reliability of the deep low-rank multi-label classification model constructed in this embodiment, the deep low-rank multi-label classification model is tested and evaluated as follows:
[0079] Input the optimal parameters obtained in step 3 into the model, use the model to perform image lesion recognition on the test set in step 1, and evaluate the model using the metrics Example-F1 and Hamming loss.
[0080] The metric Example-F1 is a combination of the average precision and recall of each sample. The larger the value of this metric, the higher the classification performance of the model. The specific calculation method is as follows:
[0081]
[0082] where p i is the precision of the sample, and r i is the recall of the sample.
[0083] The metric Hamming loss directly calculates the number of misclassified labels and is used to evaluate the average error rate of all labels. The larger the value of this metric, the lower the classification performance of the model. The specific calculation method is as follows:
[0084]
[0085] where f(x i ) is the predicted label set of the multi-label classification of sample x i , Y i is the label set, and Δ is the symmetric difference set operation.
[0086] Through the calculation of the evaluation metrics Example-F1 and Hamming loss, it is shown that the method of this embodiment meets the requirements of practical applications.
[0087] The described embodiment is a preferred embodiment of the present invention, but the present invention is not limited to the above embodiment. Without departing from the essence of the present invention, any obvious improvements, substitutions, or variations that those skilled in the art can make all fall within the protection scope of the present invention.
Claims
1. A medical lesion recognition method for deep low-rank multi-label classification, characterized in that: Collect medical lesion image data to form a sample data set, and divide the data set into a training set and a test set; construct a deep low-rank multi-label classification model based on clustering structure injection, and the deep low-rank multi-label classification model enhances the data representation ability of the original samples in the classifier through two parts: neighbor samples and low-rank representations; use the training set to train the deep low-rank multi-label classification model; use the trained deep low-rank multi-label classification model for medical lesion recognition.
2. The medical lesion recognition method for deep low-rank multi-label classification according to claim 1, wherein, Each image in the data set is a sample, represented by a feature vector.
3. The medical lesion recognition method for deep low-rank multi-label classification according to claim 2, wherein The division ratio of the training set to the test set is 6:
4.
4. The medical lesion recognition method for deep low-rank multi-label classification according to claim 1, characterized in that The construction process of the deep low-rank multi-label classification model is as follows: First, use a linear regression model as the basic model of the multi-label classification model; second, construct neighbor samples to enhance the original samples through a self-expressive subspace clustering network, and at the same time add the self-expression loss in the self-expressive subspace clustering network to the linear regression model to obtain an objective function; then, introduce low-rank representations to learn the low-rank space in the self-expressive subspace clustering network, and establish constraints to build the connection between the traditional classifier and the low-rank representations, thereby updating the objective function; finally, perform regularization constraints on the updated objective function to construct the final deep low-rank multi-label classification model.
5. The medical lesion recognition method for deep low-rank multi-label classification according to claim 4, wherein The self-expressive subspace clustering network is obtained by combining the self-expressive property and the subspace clustering network.
6. The medical lesion recognition method for deep low-rank multi-label classification according to claim 4, characterized in that The formula of the objective function is: Among them, is the dataset reconstructed using neighboring samples, w is the parameter to be learned, Y is the label, ||·|| F is the F-norm, λ1 is the hyperparameter, X is the sample dataset, θ s is the parameter of the network self-expression layer, I represents the identity matrix, and γ is the weight coefficient.
7. The medical lesion recognition method for deep low-rank multi-label classification according to claim 4, characterized in that, The formula of the updated objective function is as follows: Among them, is the dataset reconstructed using neighboring samples, w is the parameter to be learned, Y is the label, ||·|| F is the F-norm, λ1 is the hyperparameter, p is the low-rank representation matrix, X is the sample dataset, θ s is the parameter of the network self-expression layer, I represents the identity matrix, γ is the weight coefficient, and q is the weight matrix for classification in the low-dimensional subspace.
8. The medical lesion recognition method for deep low-rank multi-label classification according to claim 4, characterized in that The formula of the final deep low-rank multi-label classification model is: Among them, is the dataset reconstructed using neighboring samples, w is the parameter to be learned, ||·|| F is the F-norm, Y is the label, λ1, λ2, λ3, and λ4 are all manually set hyperparameters, p is the low-rank representation matrix, q is the weight matrix for classification in the low-dimensional subspace, X is the sample dataset, θ s is the parameter of the network self-expression layer, γ is the weight coefficient, and I is the identity matrix.
9. The method for identifying medical lesions by deep low-rank multi-label classification according to claim 8, characterized in that The training parameters of the multi-label classification model include the parameter θ of the network self-expression layer s , the parameter w to be learned, the low-rank representation matrix p, and the weight matrix q for classification in the low-dimensional subspace.
10. The medical lesion recognition method for deep low-rank multi-label classification according to claim 8, wherein The evaluation metrics of the multi-label classification model include Example-F1 and Hamming loss, and the formulas are as follows: Among them, p i is the precision of the sample, r i is the recall rate of the sample, f(x i ) is the predicted label set of the multi-label classification of the instance x i , Y i is the label set, and Δ is the symmetric difference set operation.