An ad subtype contrast clustering method based on graph-guided multi-modal decoupling
By constructing a modal adjacency matrix and a graph convolutional network through a graph-guided multimodal decoupling method, and combining an autoencoder and contrastive learning, the problem of insufficient feature differentiation between modalities in AD subtype clustering is solved, achieving higher clustering accuracy and interpretability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANTONG UNIV
- Filing Date
- 2026-02-13
- Publication Date
- 2026-06-26
Smart Images
Figure CN122290944A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of medical image processing and machine learning technology, and in particular to an AD subtype contrastive clustering method based on graph-guided multimodal decoupling. Background Technology
[0002] Alzheimer's disease (AD) is one of the leading causes of dementia worldwide. With an aging population, the number of AD cases continues to increase, placing enormous pressure on healthcare systems and significantly increasing mortality rates among the elderly. Therefore, early identification and accurate diagnosis of AD have become urgent problems to be solved. AD exhibits significant phenotypic heterogeneity, with different patients showing different progression patterns in terms of structural atrophy, functional network disruption, protein deposition, metabolic decline, and cognitive impairment. To identify subtypes of different pathological phenotypes, studies often utilize multimodal data from imaging modalities such as structural MRI, functional MRI, biomarkers (CSF), and cognitive scores. Currently, determining AD subtypes still relies heavily on the physician's personal experience and judgment, which is time-consuming and labor-intensive, and cannot adequately meet the needs of patients.
[0003] Currently, for multimodal Alzheimer's disease (AD) subtype clustering, ouzaixin et al. proposed a graph-embedded latent space learning and clustering framework in "A Graph-Embedded LatentSpace Learning and Clustering Framework for Incomplete Multimodal Multiclass Alzheimer's Disease Diagnosis." This method addresses the problem of missing multimodal data by constructing a graph-embedded latent space, projecting subjects into a low-dimensional latent representation. It combines multimodal reconstruction, subject similarity graph embedding, and AD-oriented latent clustering modules to achieve multi-class diagnosis and subtype clustering of Alzheimer's patients. Experiments show that this method performs well when handling incomplete multimodal data, effectively integrating information from different modalities and improving diagnostic accuracy. However, existing AD multimodal subtype clustering methods generally suffer from several problems, such as failing to fully utilize the neighborhood structure and biological topological relationships within modalities, making features susceptible to noise, lacking the ability to distinguish between consistent pathological information and unique pathological features between modalities, difficulty in capturing differences between subtypes, and potential inconsistencies in clustering judgments for the same patient across different modalities, thus failing to fully utilize conflicting information between modalities. Summary of the Invention
[0004] The purpose of this invention is to provide an AD subtype comparison clustering method based on graph-guided multimodal decoupling, which obtains richer pathological information of multimodal AD, improves the accuracy of subtype clustering, provides doctors with a basis for more accurate diagnosis and effectively saves time.
[0005] To achieve the aforementioned objectives, the present invention employs the following technical solution: an AD subtype contrastive clustering method based on graph-guided multimodal decoupling, comprising the following steps:
[0006] S1. Construct an affinity matrix for the original features of each modality of AD, and then obtain the adjacency matrix of each modality. Input the original features and adjacency matrix of each modality into a simple graph convolutional network to obtain the structurally enhanced features.
[0007] S2. Using the structurally enhanced features as input, an autoencoder is used to obtain latent features, which are then constrained by a reconstruction loss. The features in the latent space are then decoupled to obtain consistent pathological information across all AD modalities and supplementary information specific to each modality.
[0008] S3. Concatenate the consistency and specific features obtained after decoupling each modality to obtain the final feature representation of each modality. Cluster the final feature representation of each modality to obtain soft labels for each cluster assignment, and add cluster consistency loss to improve high-confidence samples;
[0009] S4. Based on the consistency differences and confidence scores of predictions from each modality, a hard sample contrast perception network is constructed to perform dynamic weighted contrast learning, thereby enhancing the differentiation of atypical AD cases.
[0010] S5. Minimize all loss functions and predict the clustering result for each sample, then output the final clustering result for Alzheimer's disease subtypes.
[0011] As an AD subtype contrastive clustering method based on graph-guided multimodal decoupling provided by the present invention, step S1 includes the following steps:
[0012] S11. To focus on adjacency relationships in multimodal scenarios and reduce erroneous associations between sample pairs, we first use a Gaussian kernel to calculate the affinity between each modality sample pair. The formula for calculating the affinity matrix is as follows:
[0013] (1)
[0014] In the formula This represents the i-th sample of the v-th modality. This represents the j-th sample of the v-th modality. For Gaussian kernel, The affinity matrix represents the v-th mode;
[0015] S12. After obtaining the affinity matrix, find the K nearest neighbors to the selected sample, and the adjacency matrix. The formula for determining this is as follows:
[0016] (2)
[0017] In the formula, This indicates whether the i-th sample and the j-th sample in the v-th modality are a neighbor pair. Represented as k neighbors;
[0018] S13, making Represented as The degree matrix, where yes The degree matrix can then be used to obtain a normalized affinity matrix, which is expressed as follows: Finally, the normalized graph Laplace matrix is obtained as follows: .in It is the identity matrix;
[0019] S14. Based on the assumption that graph signals should exhibit smoothness and that adjacent samples should have higher similarity, filtering graphs through graph convolutional layers is beneficial for smoothing the feature representation of clustering. However, as the number of graph convolutional layers increases, graph convolution carries the risk of over-smoothing, leading to stationary distributions in the graph signal. To mitigate over-smoothing during graph filtering, non-smooth feature graph convolution is introduced into the filtering process. The specific calculation formula is as follows:
[0020] (3)
[0021] In the formula, This represents the structure matrix after simple graph convolution. Used to balance the self-information of a node and its neighboring regions. This represents the learnable parameters shared by all views in the network. This graph operation allows for more effective capture of local correlations between different samples, contributing to the graph embedding representation of each modality.
[0022] As a graph-guided multimodal decoupling-based AD subtype contrastive clustering method provided by the present invention, step S2 includes the following steps:
[0023] S21. The features enhanced by each modal structure are used as input to the autoencoder to obtain the latent features of each modality. The autoencoder can filter out some redundant information in the modality and reduce the occurrence of the curse of dimensionality. The reconstruction process of the autoencoder is as follows:
[0024] (4)
[0025] (5)
[0026] In the formula, This represents the latent features obtained by the encoder. Indicates encoder, Indicates decoder, The neural network parameters of the encoder, This represents the neural network parameters of the decoder. This represents the reconstructed feature of the v-th mode;
[0027] S22. After reconstructing each modal feature using an autoencoder, we obtain all the reconstruction losses and need to minimize them. The calculation formula is as follows:
[0028] (6)
[0029] In the formula, Represents the reconstruction loss across all modes of Alzheimer's disease;
[0030] S23. To reduce information conflicts arising from consistent and specific features between modalities during multimodal clustering, a decoupling method is proposed. This method isolates the consistent pathological features and specific pathological information of each modality and imposes constraints to ensure complete separation. The decoupling calculation formula is as follows:
[0031] (7)
[0032] (8)
[0033] In the formula, This indicates the consistent pathological features revealed by decoupling. This represents the specific pathological information that has been decoupled. This represents a multilayer perceptron (MLP) used for modal consistency decoupling in the v-th mode. This represents a multilayer perceptron (MLP) used for modality-specific decoupling of the v-th mode.
[0034] S24. Contrastive learning is used to achieve feature consistency between views, enabling consistent features to focus on learning the shared semantics among all views. The purpose of contrastive learning is also to maximize the similarity between positive feature pairs and minimize the similarity between negative feature pairs. The calculation formula is as follows:
[0035] (9)
[0036] In the formula, The cosine similarity between two features is represented by the following expression: The temperature parameter, used for comparative learning, is used to control the smoothness of the distribution.
[0037] S25. To enhance the consistency of pathological information and the uniqueness of pathological features decoupled from the same modality, ensuring they do not contain each other's information and thus obtaining more interpretable and thoroughly decoupled features, and to prevent similarity in unique features between different modalities, thereby avoiding the contamination of the decoupling structure by false shared information, certain independence constraints are applied. The calculation formula is as follows:
[0038] (10)
[0039] In the formula, HSIC is a tool that uses kernel methods to evaluate whether two variables are independent. It can capture arbitrary nonlinear dependencies and is computationally efficient. HSIC is represented as... K is a pair The kernel matrix, L is the pair The kernel matrix, As a centered matrix, the smaller the HSIC, the more independent the features decoupled from different modes are.
[0040] As an AD subtype contrastive clustering method based on graph-guided multimodal decoupling provided by the present invention, step S3 includes the following steps:
[0041] S31. Concatenate the consistent pathological information and specific pathological features after decoupling each modality to form the final feature representation of each modality, as shown in the expression:
[0042] (11)
[0043] In the formula, This indicates the consistency and specificity features of each modality being spliced together;
[0044] S32. To obtain clustering results, a multilayer perceptron (MLP) is introduced as the clustering layer to store features. Mapping to clustering distribution In features The K-means clustering algorithm is applied to obtain pseudo-labels. To ensure consistency between the two cluster assignments, a label contrastive loss was used, calculated as follows:
[0045] (12)
[0046] In the formula, This is represented as the probability of the kth cluster.
[0047] As an AD subtype contrastive clustering method based on graph-guided multimodal decoupling provided by the present invention, step S4 includes the following steps:
[0048] S41. For the same sample i, aggregate the cluster assignments of all views:
[0049] (13)
[0050] It can be viewed as a consensus soft clustering label in multi-view semantics;
[0051] S42. For the same patient, modalities may contradict each other, resulting in large discrepancies in cluster assignments. These can be considered hard samples. First, define the confidence level within a modality as... Maximum component:
[0052] (14)
[0053] The larger the value, the more confident the modality is in clustering sample i;
[0054] S43. Next, compare the cross-modal consistency. and Differences:
[0055] (15)
[0056] The larger the value, the greater the difference between the decision of this modality and the consensus of the multimodal model, and the more difficult it is to determine the sample.
[0057] S44. Sample hardness can be determined by combining the above two weights, treating modal built-in reliability as a reliability factor (its weight should be lower if it is lower), and treating cross-modal inconsistency as another factor (its weight should be increased if it is higher). The final dynamic weight of hard samples can be expressed as:
[0058] (16)
[0059] The weights of sample i in modality v are used to dynamically identify hard samples in the modality and assign different weights to hard samples and easy samples.
[0060] S45, Cross-modal hard-sample contrast loss:
[0061] (17)
[0062] As an AD subtype contrastive clustering method based on graph-guided multimodal decoupling provided by the present invention, step S5 includes the following steps:
[0063] S51. Calculate and minimize the sum of all losses:
[0064] (18)
[0065] S52. Output the results of AD patient subtype clustering:
[0066] (19)
[0067] This is the subtype label of the AD patients. This step allows us to determine which AD subtype each patient belongs to.
[0068] Meanwhile, the present invention proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed, it implements the steps of the method described in the present invention.
[0069] Furthermore, the present invention proposes a computer-readable storage medium having a computer program stored thereon, the computer program being configured to implement the steps of the method described in the present invention when invoked by a processor.
[0070] Finally, the present invention provides a computer program product comprising a computer program / instructions that, when executed by a processor, implement the steps of the method described in the present invention.
[0071] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0072] (1) The present invention provides an AD subtype contrastive clustering method based on graph-guided multimodal decoupling, which constructs a neighborhood graph for each modality and uses graph convolution to extract structural enhancement features. At the same time, by reducing the risk of over-smoothing graph convolution, it can more effectively capture the local correlation of samples.
[0073] (2) In this invention, the embedded representation in the autoencoder latent space is decomposed into cross-modal consistent pathological information and modality-specific supplementary information, and the separation is made more complete through constraints. At the same time, cross-view contrastive learning is used to promote the alignment of consistent features, and HSIC independence constraints are used to suppress information contamination, so that consistent information and specific information do not interfere with each other. This makes up for the shortcomings of existing methods in distinguishing between intermodal consistency and specific pathological features and in capturing subtype differences, making subtype differences clearer and more interpretable.
[0074] (3) This invention proposes a hard sample comparison perception method, which first uses the built-in reliability of the modality to measure reliability, and then uses the difference in multimodal consensus allocation to measure difficulty, thereby forming dynamic weights. This allows atypical cases to receive more attention during training, and strengthens the ability to distinguish between different cases through dynamic weighted comparison learning. This specifically solves the problem that existing methods may have contradictory judgments on the same patient by different modalities and fail to make full use of conflicting information. Attached Figure Description
[0075] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.
[0076] Figure 1 This is a schematic diagram of the AD subtype comparison clustering method based on graph-guided multimodal decoupling according to the present invention.
[0077] Figure 2 This is a diagram illustrating the overall data processing framework of the method of the present invention.
[0078] Figure 3 This is a flowchart of the method of the present invention.
[0079] Figure 4 This is a framework diagram of the present invention. Detailed Implementation
[0080] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. Of course, the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0081] Example 1: As Figure 1-4 As shown: This embodiment provides an AD subtype contrastive clustering method based on graph-guided multimodal decoupling, including the following steps:
[0082] S1. Construct an affinity matrix for the original features of each modality of AD, and then obtain the adjacency matrix of each modality. Input the original features and adjacency matrix of each modality into a simple graph convolutional network to obtain the structurally enhanced features.
[0083] S2. Using the structurally enhanced features as input, an autoencoder is used to obtain latent features, which are then constrained by a reconstruction loss. The features in the latent space are then decoupled to obtain consistent pathological information across all AD modalities and supplementary information specific to each modality.
[0084] S3. Concatenate the consistency and specific features obtained after decoupling each modality to obtain the final feature representation of each modality. Cluster the final feature representation of each modality to obtain soft labels for each cluster assignment, and add cluster consistency loss to improve high-confidence samples;
[0085] S4. Based on the consistency differences and confidence scores of predictions from each modality, a hard sample contrast perception network is constructed to perform dynamic weighted contrast learning, thereby enhancing the differentiation of atypical AD cases.
[0086] S5. Minimize all loss functions and predict the clustering result for each sample, then output the final clustering result for Alzheimer's disease subtypes.
[0087] Step S1 includes the following steps:
[0088] S11. To focus on adjacency relationships in multimodal scenarios and reduce erroneous associations between sample pairs, we first use a Gaussian kernel to calculate the affinity between each modality sample pair. The formula for calculating the affinity matrix is as follows:
[0089] (1)
[0090] In the formula This represents the i-th sample of the v-th modality. This represents the j-th sample of the v-th modality. For Gaussian kernel, The affinity matrix represents the v-th mode;
[0091] S12. After obtaining the affinity matrix, find the K nearest neighbors to the selected sample, and the adjacency matrix. The formula for determining this is as follows:
[0092] (2)
[0093] In the formula, This indicates whether the i-th sample and the j-th sample in the v-th modality are a neighbor pair. Represented as k neighbors;
[0094] S13, making Represented as The degree matrix, where yes The degree matrix can then be used to obtain a normalized affinity matrix, which is expressed as follows: Finally, the normalized graph Laplace matrix is obtained as follows: .in It is the identity matrix;
[0095] S14. Based on the assumption that graph signals should exhibit smoothness and that adjacent samples should have higher similarity, filtering graphs through graph convolutional layers is beneficial for smoothing the feature representation of clustering. However, as the number of graph convolutional layers increases, graph convolution carries the risk of over-smoothing, leading to stationary distributions in the graph signal. To mitigate over-smoothing during graph filtering, non-smooth feature graph convolution is introduced into the filtering process. The specific calculation formula is as follows:
[0096] (3)
[0097] In the formula, This represents the structure matrix after simple graph convolution. Used to balance the self-information of a node and its neighboring regions. This represents the learnable parameters shared by all views in the network. This graph operation allows for more effective capture of local correlations between different samples, contributing to the graph embedding representation of each modality.
[0098] Step S2 includes the following steps:
[0099] S21. The features enhanced by each modal structure are used as input to the autoencoder to obtain the latent features of each modality. The autoencoder can filter out some redundant information in the modality and reduce the occurrence of the curse of dimensionality. The reconstruction process of the autoencoder is as follows:
[0100] (4)
[0101] (5)
[0102] In the formula, This represents the latent features obtained by the encoder. Indicates encoder, Indicates decoder, The neural network parameters of the encoder, This represents the neural network parameters of the decoder. This represents the reconstructed feature of the v-th mode;
[0103] S22. After reconstructing each modal feature using an autoencoder, we obtain all the reconstruction losses and need to minimize them. The calculation formula is as follows:
[0104] (6)
[0105] In the formula, Represents the reconstruction loss across all modes of Alzheimer's disease;
[0106] S23. To reduce information conflicts arising from consistent and specific features between modalities during multimodal clustering, a decoupling method is proposed. This method isolates the consistent pathological features and specific pathological information of each modality and imposes constraints to ensure complete separation. The decoupling calculation formula is as follows:
[0107] (7)
[0108] (8)
[0109] In the formula, This indicates the consistent pathological features revealed by decoupling. This represents the specific pathological information that has been decoupled. This represents a multilayer perceptron (MLP) used for modal consistency decoupling in the v-th mode. This represents a multilayer perceptron (MLP) used for modality-specific decoupling of the v-th mode.
[0110] S24. Contrastive learning is used to achieve feature consistency between views, enabling consistent features to focus on learning the shared semantics among all views. The purpose of contrastive learning is also to maximize the similarity between positive feature pairs and minimize the similarity between negative feature pairs. The calculation formula is as follows:
[0111] (9)
[0112] In the formula, The cosine similarity between two features is represented by the following expression: The temperature parameter, used for comparative learning, is used to control the smoothness of the distribution.
[0113] S25. To enhance the consistency of pathological information and the uniqueness of pathological features decoupled from the same modality, ensuring they do not contain each other's information and thus obtaining more interpretable and thoroughly decoupled features, and to prevent similarity in unique features between different modalities, thereby avoiding the contamination of the decoupling structure by false shared information, certain independence constraints are applied. The calculation formula is as follows:
[0114] (10)
[0115] In the formula, HSIC is a tool that uses kernel methods to evaluate whether two variables are independent. It can capture arbitrary nonlinear dependencies and is computationally efficient. HSIC is represented as... K is a pair The kernel matrix, L is the pair The kernel matrix, As a centered matrix, the smaller the HSIC, the more independent the features decoupled from different modes are.
[0116] Step S3 includes the following steps:
[0117] S31. Concatenate the consistent pathological information and specific pathological features after decoupling each modality to form the final feature representation of each modality, as shown in the expression:
[0118] (11)
[0119] In the formula, This indicates the consistency and specificity features of each modality being spliced together;
[0120] S32. To obtain clustering results, a multilayer perceptron (MLP) is introduced as the clustering layer to store features. Mapping to clustering distribution In features The K-means clustering algorithm is applied to obtain pseudo-labels. To ensure consistency between the two cluster assignments, a label contrastive loss was used, calculated as follows:
[0121] (12)
[0122] In the formula, This is represented as the probability of the kth cluster.
[0123] Step S4 includes the following steps:
[0124] S41. For the same sample i, aggregate the cluster assignments of all views:
[0125] (13)
[0126] It can be viewed as a consensus soft clustering label in multi-view semantics;
[0127] S42. For the same patient, modalities may contradict each other, resulting in large discrepancies in cluster assignments. These can be considered hard samples. First, define the confidence level within a modality as... Maximum component:
[0128] (14)
[0129] The larger the value, the more confident the modality is in clustering sample i;
[0130] S43. Next, compare the cross-modal consistency. and Differences:
[0131] (15)
[0132] The larger the value, the greater the difference between the decision of this modality and the consensus of the multimodal model, and the more difficult it is to determine the sample.
[0133] S44. Sample hardness can be determined by combining the above two weights, treating modal built-in reliability as a reliability factor (its weight should be lower if it is lower), and treating cross-modal inconsistency as another factor (its weight should be increased if it is higher). The final dynamic weight of hard samples can be expressed as:
[0134] (16)
[0135] The weights of sample i in modality v are used to dynamically identify hard samples in the modality and assign different weights to hard samples and easy samples.
[0136] S45, Cross-modal hard-sample contrast loss:
[0137] (17)
[0138] Step S5 includes the following steps:
[0139] S51. Calculate and minimize the sum of all losses:
[0140] (18)
[0141] S52. Output the results of AD patient subtype clustering:
[0142] (19)
[0143] This is the subtype label of the AD patients. This step allows us to determine which AD subtype each patient belongs to.
[0144] Example 2: In this example, the multimodal data includes any two or more combinations of the following: structural magnetic resonance imaging data, functional magnetic resonance imaging data, diffusion tensor imaging data, clinical scale data, and genetic biomarker data. Corresponding modal graph structures are constructed for each modal data, and feature enhancement and other operations are performed according to the steps described in Example 1.
[0145] Example 3: In this example, the number of nearest neighbors K selected when constructing the adjacency matrix is a preset positive integer, which can be adjusted according to the sample size and modal characteristics. During graph convolution, the weight parameters used to balance node self-information and neighborhood information in non-smooth graph convolution are learnable parameters or pre-set hyperparameters. In contrastive learning, the temperature parameter is a positive real number, and its value range can be adjusted according to training stability.
[0146] Example 4: Based on Examples 1-3, this example expands the application scenarios of the method. In this example, the method is not only applicable to subtype clustering of Alzheimer's patients, but can also be applied to subtype analysis of other mental illnesses, patient classification of other neurodegenerative diseases, and unsupervised clustering analysis of multimodal medical data.
[0147] Example 5: This example provides a multimodal Alzheimer's patient subtype clustering system for implementing the above method. For example... Figure 4 As shown, it includes modules for data acquisition, graph structure construction, feature enhancement, representation learning and decoupling, hard sample mining, and result output.
[0148] Example 6: This example provides a computer device, including a memory and a processor. When executed by the processor, the computer program implements the method steps as described in any one of Examples 1 to 5. Furthermore, this example also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described multimodal Alzheimer's disease patient subtype clustering method.
[0149] The program code used to implement the methods of this application may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.
[0150] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. An AD subtype contrast clustering method based on graph-guided multi-modal decoupling, characterized in that, Includes the following steps: Step S1: Construct an affinity matrix for the original features of each modality of Alzheimer's disease (AD) to obtain the adjacency matrix of each modality. Input the original features and adjacency matrix of each modality into a simple graph convolutional network to obtain structurally enhanced features. Step S2: The structurally enhanced features are used as input, and the latent features are obtained through an autoencoder. The features in the latent space are decoupled using the reconstruction loss constraint to obtain consistent pathological information for each modality of Alzheimer's disease (AD) and supplementary information unique to each modality. Step S3: Concatenate the consistency features and specific features obtained after decoupling each modality to obtain the final feature representation of each modality. Cluster the final feature representation of each modality to obtain the cluster assignment soft label for each modality, and add cluster consistency loss to improve high confidence samples. Step S4: Construct a hard sample contrast perception network based on the consistency differences and confidence scores of each modality prediction to perform dynamic weighted contrast learning, thereby enhancing the differentiation of atypical Alzheimer's disease (AD) cases; Step S5: Minimize all loss functions and predict the clustering result for each sample, then output the final clustering result for Alzheimer's disease subtypes.
2. The Alzheimer's disease (AD) subtype contrast clustering method based on graph-guided multi-modal decoupling according to claim 1, wherein, Step S1 includes the following steps: S11. Calculate the affinity between each modality sample pair using a Gaussian kernel. The formula for calculating the affinity matrix is as follows: （1）； wherein represents the i-th sample of the v-th modality, represents the j-th sample of the v-th modality, is a Gaussian kernel, represents the affinity matrix of the v-th modality; S12. After obtaining the affinity matrix, find the K nearest neighbors to the selected sample, and the adjacency matrix. The formula for determining this is as follows: （2）； In the formula, This indicates whether the i-th sample and the j-th sample in the v-th modality are neighbor pairs. Represented as k neighbors; S13, making Represented as The degree matrix, where yes The degree matrix is used to obtain a normalized affinity matrix, which is expressed as follows: Finally, the normalized graph Laplace matrix is obtained as follows: ,in It is the identity matrix; S14. Based on the assumption that graphical signals should exhibit smoothness and that adjacent samples are similar, graph convolutional layers are used to filter the graph. However, as the number of graph convolutional layers increases, there is a risk of over-smoothing, leading to a stationary distribution in the graph signal. Therefore, non-smooth feature graph convolution is introduced into the filtering process, and the calculation formula is as follows: （3）； In the formula, This represents the structure matrix after simple graph convolution. Used to balance the self-information of a node and its neighboring regions. This represents the learnable parameters shared by all views in the network.
3. The Alzheimer's disease (AD) subtype contrastive clustering method based on graph-guided multimodal decoupling according to claim 1, characterized in that, Step S2 includes the following steps: S21. The features enhanced by each modal structure are used as input to the autoencoder to obtain the latent features of each modality. The autoencoder then filters out some redundant information in the modality. The reconstruction process of the autoencoder is as follows: （4）； （5）； In the formula, This represents the latent features obtained by the encoder. Indicates encoder, Indicates decoder, The neural network parameters of the encoder, This represents the neural network parameters of the decoder. This represents the reconstructed feature of the v-th mode; S22. After reconstructing each modal feature using an autoencoder, obtain all reconstruction losses and minimize them. The calculation formula is as follows: （6）； In the formula, Represents the reconstruction loss across all modes of Alzheimer's disease; S23. Identify the consistent pathological features and specific pathological information of each modality independently, and impose constraints to ensure complete separation. The decoupling calculation formula is as follows: （7）； （8）； In the formula, This indicates the consistent pathological features revealed by decoupling. This represents the specific pathological information that has been decoupled. This represents a multilayer perceptron (MLP) used for modal consistency decoupling in the v-th mode. This represents a multilayer perceptron (MLP) used for modality-specific decoupling of the v-th mode; S24. A contrastive learning method is used to achieve feature consistency between views, so that consistent features focus on learning the shared semantics among all views. The purpose of contrastive learning is to maximize the similarity between positive feature pairs and minimize the similarity between negative feature pairs. The calculation formula is as follows: （9）； In the formula, The cosine similarity between two features is represented by the following expression: The temperature parameter, represented in the comparative learning, is used to control the smoothness of the distribution; S25. Apply independence constraints, the calculation formula is as follows: （10）； The Hilbert-Schmidt Independence Criterion (HSIC) is a tool that uses a kernel method to assess whether two variables are independent. HSIC is expressed as... K is a pair The kernel matrix, L is the pair The kernel matrix, As a centered matrix, the smaller the HSIC, the more independent the features decoupled from different modes are.
4. The AD subtype contrastive clustering method based on graph-guided multimodal decoupling according to claim 1, characterized in that, Step S30 includes the following steps: S31. Concatenate the consistent pathological information and specific pathological features after decoupling each modality to form the final feature representation of each modality, as shown in the expression: （11）； In the formula, This indicates the consistency and specificity features of each modality being spliced together; S32. To obtain clustering results, a multilayer perceptron (MLP) is introduced as the clustering layer to store features. Mapping to cluster distribution In features The K-means clustering algorithm is applied to obtain pseudo-labels. To ensure consistency in the two cluster assignments, label contrast loss was used, calculated as follows: （12）； In the formula, This is represented as the probability of the kth cluster.
5. The AD subtype contrastive clustering method based on graph-guided multimodal decoupling according to claim 1, characterized in that, Step S40 includes the following steps: S41. For the same sample i, aggregate the cluster assignments of all views: （13）； For consensus soft clustering labels in multi-view semantics; S42. For the same patient, modalities may contradict each other, resulting in large discrepancies in cluster assignments. These are considered hard samples. First, the confidence level within a modality is defined as... Maximum component: （14）； The larger the value, the more confident the modality is in clustering sample i; S43. Next, compare the cross-modal consistency. and Differences: （15）； S44. Treating cross-modal inconsistency as another factor, the final hard sample dynamic weights are expressed as: （16）； Assign weights to sample i in modality v, dynamically identify hard samples in the modality, and assign different weights to hard samples and easy samples; S45, Cross-modal hard-sample contrast loss: （17）。 6. The AD subtype contrastive clustering method based on graph-guided multimodal decoupling according to claim 1, characterized in that, Step S50 includes the following steps: S51. Calculate and minimize the sum of all losses: （18）； S52. Output the results of Alzheimer's disease (AD) patient subtype clustering: （19）； This outputs the subtype label of Alzheimer's disease (AD) patients. This step determines which subtype of Alzheimer's disease each patient belongs to.
7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed, it implements the steps of the method as described in any one of claims 1 to 6.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is configured to implement the steps of the method according to any one of claims 1 to 6 when invoked by a processor.
9. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 6.