Robust bipartite graph medical image feature learning method based on correlation entropy

Abstract

The invention discloses a medical image feature learning method for a robust bipartite graph based on correlation entropy, and belongs to the field of computer vision and image analysis, and the method mainly comprises the steps: generating an anchor point set through employing a non-negative linear combination method; constructing an adjacent matrix between the sample and the anchor points based on the anchor point set, and forming a bipartite graph structure; on the basis of the maximum correlation entropy criterion, feature reconstruction is carried out on the sample data feature matrix in an anchor point weighted combination mode; constructing a joint optimization model, optimizing the feature expression in the graph embedding space through an iteration mode, and when the joint optimization model converges, outputting the low-dimensional embedding expression of the sample; and based on the low-dimensional embedding representation, a final clustering result is obtained by adopting a clustering algorithm. The method not only has the end-to-end graph structure and feature joint learning ability, but also considers the model generalization ability and the calculation efficiency, and is especially suitable for performing semi-supervised feature learning and clustering tasks in medical image scenes with artifacts, uneven distribution or scarce marks.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and image analysis, and more particularly to a robust bipartite graph-based method for learning medical image features. It can be used to improve the robustness of image clustering and recognition under outlier interference conditions, and is especially suitable for semi-supervised scenarios in medical images such as MRI, CT, or pathology images where artifacts, missing, or anomalous samples exist. Background Technology

[0002] Image clustering is a fundamental task in medical image analysis, with wide-ranging value in applications such as disease classification, lesion identification, case retrieval, and assisted diagnosis. With the development of anchor graph modeling and graph representation learning techniques, bipartite graph structures, due to their ability to effectively describe the low-rank adjacency relationships between samples and anchor points, have been increasingly used in large-scale image clustering. However, in actual medical image acquisition and processing, medical images are often affected by factors such as noise, occlusion, motion artifacts, and equipment differences, resulting in a large number of outliers or abnormal samples, which in turn affect the accuracy and stability of clustering. Especially in semi-supervised environments, the high cost and sparse annotation of existing methods make them even more susceptible to the influence of outliers.

[0003] Some graph clustering methods have attempted to improve the robustness of the model by introducing strategies such as graph regularization and adaptive weight learning, but the following problems still exist: 1. Sensitive to outliers: Traditional graph construction methods based on Euclidean distance or mean squared error lack robustness to outliers; 2. Fixed structure: Most methods use predefined graph structures that cannot adaptively adjust to data distribution; 3. High computational complexity: Some methods require multiple iterations or solving complex optimization problems, making them difficult to apply to large-scale data; 4. Insufficient use of labels: Semi-supervised information was not effectively integrated into the graph construction and feature learning process.

[0004] While the Maximum Correlation Entropy (MCC) criterion has been proven effective in suppressing the influence of outliers in robust regression and classification tasks, its application in graph clustering, especially bipartite graph modeling, has not been fully explored. Therefore, existing techniques still face challenges such as poor robustness, weak generalization ability, and high computational cost in processing noisy, outlier, and poorly labeled medical image clustering tasks. There is an urgent need for an efficient medical image learning method that integrates robust information theory metrics and adjustable graph structure mechanisms to improve the performance of medical image clustering in complex environments. Summary of the Invention

[0005] The purpose of this invention is to overcome the problems existing in the prior art and provide a robust bipartite graph-based method for learning medical image features based on correlation entropy.

[0006] The objective of this invention is achieved through the following technical solution: A robust bipartite graph-based method for learning medical image features, comprising the following steps: S1. Input the original medical image sample, extract the features of the original medical image sample, and obtain the sample data feature matrix; S2. Based on the feature matrix of the sample data, an anchor point set is generated using a non-negative linear combination method; S3. Construct an adjacency matrix between samples and anchor points based on the anchor point set, and form a bipartite graph structure; S4. Based on the maximum correlation entropy criterion, the feature matrix of the sample data is reconstructed by weighted combination of anchor points; S5. Construct a joint optimization model containing multiple variables, optimize the feature representation in the graph embedding space through an iterative method, and output the low-dimensional embedding representation of the sample when the joint optimization model converges; S6. Based on the low-dimensional embedding representation, a clustering algorithm is used to obtain the final clustering result.

[0007] Preferably, the method of generating the anchor point set using non-negative linear combination includes: Constructing the anchor matrix ,in The weight matrix represents the contribution of a sample to each anchor point; It satisfies the nonnegativity and normalization constraints.

[0008] Preferably, step S3 specifically includes: Calculate the Euclidean distance between the anchor point and the sample; Construct an adjacency matrix based on the Euclidean distance between the anchor point and the sample. Adjacency matrix The elements in are defined as: in, This represents a sample in the feature matrix of the sample data. Indicates bandwidth parameter, Indicates the anchor point, when the anchor point Belongs to the sample It is valid when the k-nearest neighbor is found, otherwise it is 0, thus forming a bipartite graph structure.

[0009] Preferably, a robust loss function is introduced in step S4, and the robust loss function includes: Relevant entropy terms: ; Tag guide items: Where Z is the structure representation matrix, and M is the low-dimensional mapping of the anchor points. Let be the i-th column of the adjacency matrix H. Let i be the i-th column of the constraint matrix A; Sparse regularization terms: ,in, Represents the principal regularization coefficient. This represents the hyperparameter used for the sparse regularization term.

[0010] Preferably, the objective function of the joint optimization model is: in, For the projection matrix, This represents the relevant entropy term.

[0011] Preferably, in the optimization process of step S5, an equivalent substitution form based on Fenchel conjugate transformation is adopted to transform the exponential loss into a weighted quadratic loss, and then an accelerated block coordinate update algorithm is applied to solve it.

[0012] Preferably, the final clustering result is obtained through one of the following methods: Perform k-means or spectral clustering on the low-dimensional embedding representation to obtain the final cluster labels; Will Clustering labels are set to The better clustering result is selected.

[0013] It should be further noted that the technical features corresponding to the above options can be combined or substituted to form new technical solutions if there is no conflict.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. The method of the present invention can effectively mitigate the impact of abnormal samples, noise interference and structural bias on clustering results without relying on a large amount of labeled data.

[0015] 2. The method of the present invention improves the clustering robustness and discrimination ability of the graph model in complex medical image scenarios by constructing an adaptively optimizable sample-anchor bipartite graph structure and introducing the maximum correlation entropy criterion to enhance the suppression of outliers.

[0016] 3. The method of this invention not only possesses end-to-end graph structure and feature joint learning capabilities, but also balances model generalization ability and computational efficiency. The method can be flexibly applied to various feature learning models, exhibiting strong adaptability. It is particularly suitable for semi-supervised feature learning and clustering tasks in medical imaging scenarios with artifacts, uneven distribution, or scarce annotations.

[0017] 4. The method of this invention achieves better clustering performance than existing methods on three publicly available image datasets, with a significant improvement in clustering accuracy. Attached Figure Description

[0018] Figure 1 This is a flowchart of a robust bipartite graph-based medical image feature learning method based on correlation entropy, according to the present invention. Figure 2 This is a schematic diagram of the network structure of the robust bipartite graph feature learning method based on relevance entropy of the present invention; Figure 3 Visualization results of different anchor point selection methods of the present invention; Figure 4 This diagram illustrates the comparison of the effects of different methods of the present invention on three datasets. Detailed Implementation

[0019] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] It should be noted that the defects in the solutions in the prior art are all the results of the inventors' practice and careful research. Therefore, the discovery process of the above problems and the solutions proposed by the embodiments of this application in the following text should be the inventors' contributions to this application in the process of invention and creation, and should not be understood as technical content known to those skilled in the art.

[0021] In one exemplary embodiment, such as Figure 1 As shown, a robust bipartite graph-based method for learning medical image features is provided, comprising the following steps: S1. Input the original medical image sample, extract the features of the original medical image sample, and obtain the sample data feature matrix; S2. Based on the feature matrix of the sample data, an anchor point set is generated using a non-negative linear combination method; S3. Construct an adjacency matrix between samples and anchor points based on the anchor point set, and form a bipartite graph structure; S4. Based on the maximum correlation entropy criterion, the feature matrix of the sample data is reconstructed by weighted combination of anchor points; S5. Construct a joint optimization model containing multiple variables, optimize the feature representation in the graph embedding space through an iterative method, and output the low-dimensional embedding representation of the sample when the joint optimization model converges; S6. Based on the low-dimensional embedding representation, a clustering algorithm is used to obtain the final clustering result.

[0022] For example, in step S1, the sample data feature matrix Each column represents a d-dimensional sample feature vector, with a total of n samples. In this invention, the samples can be depth features of the original image, PCA dimensionality reduction results, or other preprocessed feature representations.

[0023] For example, in step S2, a set of anchor points is generated using a non-negative linear combination method, and an anchor point matrix is ​​constructed. ,in Let W be the weight matrix, representing the contribution of a sample to each anchor point. This weight matrix W must satisfy non-negativity and normalization constraints, i.e. ,and This ensures that each anchor point is a weighted combination of all samples.

[0024] For example, in step S3, based on the Euclidean distance between the anchor point and the sample, the adjacency matrix of the sample and anchor point is further constructed. ,in When anchor point Belongs to the sample The adjacency matrix is ​​valid only when the k nearest neighbor is found, otherwise it is 0, thus forming a bipartite graph structure. This adjacency matrix is ​​used for subsequent feature propagation and graph modeling.

[0025] For example, in step S4, feature reconstruction and robust loss are introduced. To improve the model's robustness to noise and outliers, this invention introduces the Maximum Relevance Entropy (MCC) criterion during sample reconstruction. Samples are reconstructed using a weighted combination of anchor points, with the goal of minimizing the difference between the sample and the reconstruction result. Unlike traditional mean squared error loss, this invention employs a correlation entropy loss function with an exponential kernel to suppress the influence of large errors, defined as follows: Simultaneously, a label-guided mechanism is introduced, which incorporates some known label information into the model by constructing a semi-supervised constraint matrix A. Specifically, the label-guided constraint terms take the form of... Where Z is the structure representation matrix and M is the low-dimensional mapping of anchor points. Let be the i-th column of the adjacency matrix H. Let be the i-th column of the constraint matrix A. The complete robust loss function is the sum of these two terms minus the sparse regularization term. This is used to enhance the selectivity of graph structures, thereby improving clustering performance and robustness.

[0026] For example, in step S5, robust graph learning joint optimization is performed. To achieve collaborative learning of graph structure and anchor representation, this invention constructs a joint optimization model containing multiple variables, aiming to maximize the fusion benefit of robust graph structure learning and low-rank embedding. The joint optimization function is in the form of: maximize ,in Let be the projection matrix, satisfying The orthogonal constraint represents the sample representation in the low-dimensional space; The aforementioned entropy loss term is represented by [reference needed]. Model parameters include anchor weights W, adjacency matrix H, structural representation matrix Z, anchor representation M, and feature map F, all of which are jointly optimized iteratively. During optimization, an equivalent substitution based on Fenchel conjugate transformation is used to convert the exponential loss into a weighted quadratic loss, which is then solved using the Accelerated Block Coordinate Update (BCU) algorithm. This algorithm performs closed-form or approximate optimization on each variable while maintaining constraints. For example, W must satisfy the normalization constraint for each column, which can be achieved by projecting onto the probabilistic simplex space; Z, M, and H must remain non-negative, achieved through non-negative projection; and the update of F is based on singular value decomposition while maintaining orthogonality. This optimization framework ensures that each update step converges to a local optimum, maintaining the monotonic improvement of the objective function and effectively improving the overall clustering accuracy. After the optimization process converges, this invention finally constructs a low-dimensional embedding representation ZA of the samples using the learned structural representation matrix Z and the semi-supervised constraint matrix A.

[0027] For example, in step S6, clustering output is performed. The low-dimensional embedding representation ZA can be used in traditional clustering methods such as k-means or spectral clustering to obtain the final cluster label, or to extract the instance... Clustering labels are set to Then, the better clustering result is selected. Each column represents a sample in a low-dimensional space, and clustering these columns yields the sample's category. Furthermore, if labels exist, performance can be evaluated by calculating metrics such as clustering accuracy (ACC) and normalized mutual information (NMI). This method organically integrates the structural information of the bipartite graph, label supervision information, and robust feature representation, possessing advantages such as strong scalability, adaptability, and robustness. It is suitable for various tasks including image clustering and sample classification, maintaining excellent performance even in the presence of noise and incomplete labels.

[0028] Furthermore, a specific implementation example of the above method is given, with the overall network structure as follows: Figure 2 As shown, starting with the original input sample X, and combining some known label information A, the robust low-dimensional feature representation Z is learned through steps such as anchor point selection, bipartite graph construction, and low-rank matrix factorization, and finally the clustering result is output.

[0029] Specifically, the original sample X contains several data points to be clustered, some of which have known label information, used to construct a label matrix A. Then, an initial feature representation W is generated through feature transformation, from which representative anchor points U are selected. Next, a bipartite graph structure between samples and anchor points is constructed, forming an adjacency relation matrix H, which describes the connection weights between samples and anchor points.

[0030] To improve the robustness of feature representation, a low-rank matrix factorization mechanism based on correlation entropy is introduced. This module includes variables such as the basis matrix F, the low-dimensional feature representation matrix Z, and the auxiliary reconstruction matrix M. By optimizing the correlation entropy objective function, it effectively resists interference from noise or uncertain labels while maintaining structural consistency. The low-rank factorization model can be represented as A≈FZ or A≈FM, and is used to guide robust feature extraction.

[0031] Furthermore, to fully utilize label information and graph structure, an adaptive feature learning mechanism is proposed. This mechanism iteratively optimizes the feature representation Z in the graph embedding space, achieving a coarse-to-fine feature representation update. Finally, based on the optimized Z and graph structure, unsupervised clustering is performed on all samples, and the clustering results are output, denoted as "Category 1" and "Category 2" in the figure.

[0032] This invention evaluates the rationality and interpretability of the optimized anchors by conducting experiments on synthetic datasets. Specifically, the final anchors were acquired and visualized, and compared with anchors generated by other anchor selection methods, including K-means and ABG-LMF without MCC. The visualization results are shown in Figure 3. It was observed that the anchors generated by the ABG-LMF method effectively capture the relationships between samples. In contrast, K-means is highly sensitive to outliers, resulting in some anchors being located at outlier positions. Meanwhile, the anchors generated by ABG-LMF without MCC are more randomly scattered and cannot adequately represent the relationships between samples and anchors.

[0033] To evaluate the stability of the proposed method, artificial outliers were introduced into the dataset by generating virtual images matching the size of the original samples. Each pixel in these virtual images was randomly assigned a value of 1 or 255. These artificial samples comprised 20% of the total data instances. To comprehensively evaluate clustering performance, experiments were repeated on three datasets (Yale, ATT, and WarpAR). Figure 4 visualizes the relationship between the number of clusters and the clustering results. Experimental results show that the ABG-LMF method consistently demonstrates good clustering performance on the Yale, ATT, and WarpAR datasets, and its average performance generally outperforms existing methods at different numbers of clusters. On the Yale dataset, the average accuracy of the ABG-LMF method is 75.51%, which is 8.67% higher than the second-ranked method CSNMF (66.84%). On the ATT dataset, the average accuracy of the ABG-LMF method is 96.79%, which is 5.7% higher than the second-ranked RDCF (91.09%). On the WarpAR dataset, the ABG-LMF method achieved an average accuracy of 68.21%, which is 11.69% higher than the second-place EAGR (56.52%).

[0034] The above detailed embodiments are a description of the present invention. It should not be considered that the specific embodiments of the present invention are limited to these descriptions. For those skilled in the art, several simple deductions and substitutions can be made without departing from the concept of the present invention, and all of these should be considered to fall within the protection scope of the present invention.

Claims

1. A robust bipartite graph-based medical image feature learning method based on correlation entropy, characterized in that, Includes the following steps: S1. Input the original medical image sample, extract the features of the original medical image sample, and obtain the sample data feature matrix; S2. Based on the feature matrix of the sample data, an anchor point set is generated using a non-negative linear combination method; S3. Construct an adjacency matrix between samples and anchor points based on the anchor point set, and form a bipartite graph structure; S4. Based on the maximum correlation entropy criterion, the feature matrix of the sample data is reconstructed by weighted combination of anchor points; S5. Construct a joint optimization model containing multiple variables, optimize the feature representation in the graph embedding space through an iterative method, and output the low-dimensional embedding representation of the sample when the joint optimization model converges; S6. Based on the low-dimensional embedding representation, a clustering algorithm is used to obtain the final clustering result.

2. The method for learning medical image features based on a robust bipartite graph with correlation entropy according to claim 1, characterized in that, The method of generating the anchor point set using non-negative linear combination includes: Constructing the anchor matrix ,in The weight matrix represents the contribution of a sample to each anchor point; It satisfies the nonnegativity and normalization constraints.

3. The method for learning medical image features based on robust bipartite graphs with correlation entropy according to claim 1, characterized in that, Step S3 specifically includes: Calculate the Euclidean distance between the anchor point and the sample; Construct an adjacency matrix based on the Euclidean distance between the anchor point and the sample. Adjacency matrix The elements in are defined as: in, This represents a sample in the feature matrix of the sample data. Indicates bandwidth parameter, Indicates the anchor point, when the anchor point Belongs to the sample It is valid when the k-nearest neighbor is found, otherwise it is 0, thus forming a bipartite graph structure.

4. The method for learning medical image features based on a robust bipartite graph with correlation entropy according to claim 3, characterized in that, In step S4, a robust loss function is introduced, which includes: Relevant entropy terms: ; Tag guide items: Where Z is the structure representation matrix, and M is the low-dimensional mapping of the anchor points. Let be the i-th column of the adjacency matrix H. Let i be the i-th column of the constraint matrix A; Sparse regularization terms: ,in, Represents the principal regularization coefficient. This represents the hyperparameter used for the sparse regularization term.

5. The method for learning medical image features based on a robust bipartite graph with correlation entropy according to claim 4, characterized in that, The objective function of the joint optimization model is: in, Let be the projection matrix. This represents the relevant entropy term.

6. The method for learning medical image features based on a robust bipartite graph with correlation entropy according to claim 5, characterized in that, In the optimization process of step S5, an equivalent substitution form based on Fenchel conjugate transformation is adopted to transform the exponential loss into a weighted quadratic loss, and then an accelerated block coordinate update algorithm is applied to solve it.

7. The method for learning medical image features based on a robust bipartite graph with correlation entropy according to claim 5, characterized in that, The final clustering results are obtained through one of the following methods: Perform k-means or spectral clustering on the low-dimensional embedding representation to obtain the final cluster labels; Will Clustering labels are set to The better clustering result is selected.

Images