A multi-modal ad staging three-branch contrast clustering method based on multi-particle co-occurrence guidance

Abstract

The application provides a multi-modal AD staging three-branch contrast clustering method based on multi-particle co-occurrence guidance, comprising the following steps: S10, acquiring multi-modal clinical data of AD patients; S20, constructing a fuzzy granular neighborhood of each modality data based on granular computing theory, realizing feature enhancement and noise suppression within the mode; S30, constructing a cross-modal gated attention fusion network, completing the deep interaction and sample-level fusion of heterogeneous features; S40, designing a confidence level hierarchical contrast learning strategy for different decision domains, and strengthening the discriminative feature learning of boundary samples; S50, outputting the staging clustering results of Alzheimer's disease patients. The application effectively improves the clustering accuracy of multi-modal data of Alzheimer's disease and the fitting degree of clinical staging, reduces the misjudgment risk of fuzzy samples in clinical staging. It provides a reliable artificial intelligence auxiliary basis for early disease screening, accurate staging and personalized diagnosis and treatment.

Description

Technical Field

[0001] This invention belongs to the field of medical data mining and artificial intelligence-assisted diagnosis technology, and in particular relates to a three-branch comparative clustering method for multimodal AD staging based on multi-sphere co-occurrence. Background Technology

[0002] Alzheimer's disease (AD), a progressive neurodegenerative disease, involves a complex and multidimensional process in its pathological evolution, encompassing brain anatomical atrophy, abnormal neural metabolic networks, and alterations in the body's humoral microenvironment. With advancements in neuroimaging and molecular biology, structural magnetic resonance imaging (sMRI), positron emission tomography (PET), cerebrospinal fluid biomarkers, and cognitive scales have created multi-source, heterogeneous datasets describing the pathological state of AD. While these multimodal data provide rich information redundancy for early identification and accurate staging of AD across different physical and physiological dimensions, they also present significant challenges due to data heterogeneity. The core issue lies in how to leverage artificial intelligence to penetrate the surface of multi-source data and establish a nonlinear mapping model from high-dimensional, heterogeneous multimodal features to the continuous pathological staging of AD, thereby achieving objective and accurate individualized diagnosis and treatment. This has become a crucial current challenge.

[0003] Currently, numerous computational methods for auxiliary diagnosis of Alzheimer's disease (AD) have emerged, but they generally suffer from two limitations: First, the one-sided utilization of modalities. Most existing mainstream methods rely on a single modality (such as morphological analysis based solely on sMRI or threshold judgment of biomarkers), which severs the intrinsic synergistic mechanisms of AD pathology at the imaging, metabolic, and molecular levels, making it difficult to comprehensively depict the heterogeneity of the disease. Second, the superficiality of fusion strategies. Some studies attempting multimodal fusion often employ feature-level splicing or decision-level weighting strategies. This linear or superficial combination approach fails to effectively model the nonlinear deep interactions and complementary information between modalities and often ignores the topological proximity relationships of sample data in the potential manifold space, resulting in insufficient discriminative power of fused features for subtle disease progression. Summary of the Invention

[0004] Purpose of the invention: The technical problem to be solved by the present invention is to provide a three-branch comparative clustering method for multimodal Alzheimer's disease staging based on the co-occurrence of multiple spheres, which addresses the shortcomings of the existing technology. This method solves the technical problems of insufficient fusion of modal features heterogeneity, low discrimination of boundary samples, and poor fit between clustering results and staging in the existing unsupervised clustering of multimodal Alzheimer's disease data.

[0005] The method of the present invention includes the following steps:

[0006] Step S10: Obtain multimodal data of AD patients and construct independent encoding and decoding module groups, and drive model learning by minimizing reconstruction loss;

[0007] Step S20: Construct fuzzy particle sphere neighborhoods for each modality based on particle sphere computation theory, capture local structural associations of samples through multi-particle sphere co-occurrence similarity measurement, and achieve intra-modal feature enhancement and noise suppression;

[0008] Step S30: Construct a cross-modal gated attention fusion network, model the complementary correlation between modalities using the cross-attention mechanism, and dynamically allocate modal contribution weights through adaptive gating units to complete the deep interaction and sample-level fusion of heterogeneous features;

[0009] Step S40: Based on the conflict correction fusion similarity matrix, the sample space is divided into positive domain, negative domain and boundary domain through three-way decision. A confidence hierarchical comparison learning strategy is designed for different decision domains to strengthen the discriminative feature learning of boundary samples.

[0010] Step S50, Joint comparison loss With reconstruction loss The model is iteratively optimized, and the final output is the staged clustering results of Alzheimer's patients.

[0011] Step S10 includes the following steps:

[0012] Step S11: Obtain multimodal data of AD patients, including structural magnetic resonance imaging and fluorodeoxyglucose positron emission tomography neuroimaging data, molecular biomarkers such as Aβ42 protein and tau protein in cerebrospinal fluid, as well as cognitive assessment scores of Mini-Mental State Examination and Alzheimer's Disease Rating Scale.

[0013] Multimodal datasets ,in This represents the total number of AD patient samples. Indicates the first A sample of AD patients, where i takes values ​​from 1 to n;

[0014] The i-th AD patient sample Represented as a multidimensional feature vector ,in Representative sample Structural Magnetic Resonance Imaging (sMRI) structural feature vectors, Representative sample Metabolic feature vectors of fluorideoxyglucose positron emission tomography (FDG-PET). Representative sample The β-amyloid positron emission tomography (Aβ-PET) eigenvectors characterize the Aβ deposition load in brain regions. Representative sample cerebrospinal fluid / blood biomarker vectors, among which, and The concentration values ​​of β-amyloid protein with 42 residues and 40 residues, respectively; This represents the total tau protein concentration. This represents the concentration of tau protein phosphorylated at threonine 217. Representative sample The apolipoprotein E (APOE) genotype vector, representing the number of ε4 alleles. Representative sample Cognitive function score vector, each component , The quantitative scores correspond sequentially to the Minimal Mental State Examination (MMSE), Montreal Cognitive Assessment (MoCA), Alzheimer's Disease Rating Scale-Summarized Box (CDR-SB), and the Alzheimer's Disease Assessment Scale-Cognitive Subscale (ADAS-Cog), and satisfy the total dimensions of the original characteristics. , Let represent the real space, the superscript indicates the dimension of the corresponding vector, and T denotes the transpose of the vector. These are the characteristic dimensions of sMRI, FDG-PET, and Aβ-PET, respectively; 3, 1, and 4 are the fixed dimensions of biomarkers, genotypes, and cognitive scores.

[0015] Step S12: Construct independent coding modules for each modality and decoder module ,in Indicates the total number of modalities and each encoding module. Mapping the original features to a latent representation of uniform dimension; each decoder module The latent representation is reconstructed into the original feature space, where For network parameters, Let m be the original feature dimension. For the unified dimensions;

[0016] The encoding process generates a low-dimensional feature set through reparameterization techniques. The definition is as follows:

[0017] (1),

[0018] in , To obtain from the multivariate standard normal distribution The noise vector sampled in the middle, This represents a normal distribution with a mean of zero and a covariance equal to the identity matrix. These are the mean and variance vectors of the output, respectively. For the sample In the Semantic vectors in each modality This is an element-wise multiplication operation;

[0019] The decoding process will represent the latent representation Reconstructing Back to the First The original feature space of each modality :

[0020] (2),

[0021] The reconstruction loss calculation uses the mean squared error to measure the difference between the original features and the reconstructed features, and minimizes the reconstruction loss. Optimize encoder parameters and decoder parameters , making the potential representation Able to preserve original modal features to the greatest extent Semantic information.

[0022] In step S12, the reconstruction loss The formula is as follows:

[0023] (3).

[0024] Step S20 includes the following steps:

[0025] Step S21, in each mode low-dimensional feature subspace In the middle, order Indicates based on The generated set of spheres, in which Indicates the total number of balls. , For granules Number of samples center and radius The definition is as follows:

[0026] (4),

[0027] First, initialize the entire dataset as a single sphere, if the partitioning condition is met. and Then, 2-means is used to split the granule into two sub-granules. and Distribution measure The definition is as follows:

[0028] (5);

[0029] Step S22: For each sample in the m-th mode ,calculate The membership degree of the top-p nearest spheres in terms of geometric distance, and the sample The membership degree of the remaining particles is set to 0; in mode m, the j-th sample For the i-th ball membership degree The definition is as follows:

[0030] (6),

[0031] in It is a hyperparameter that adjusts the degree of blur. It is a sample of Nearest neighbor set of particles, membership matrix exp represents the natural exponential function;

[0032] Step S23: Calculate the fuzzy co-occurrence similarity between samples to construct a pathological adjacency graph: For any two samples and ,if and For the same ball membership degree and If all values ​​are not less than the preset threshold η, it indicates that the value is within the range of... Patients are defined as being at similar stages of disease progression within a localized pathological space at a given scale. With patients Connection weights between For the fuzzy Jaccard similarity on all spheres:

[0033] (7),

[0034] in, It is a smoothing term to avoid a denominator of 0; graph adjacency matrix. ;

[0035] Step S24, for Normalization is performed to obtain the normalized Laplace matrix. :

[0036] (8),

[0037] in It is the identity matrix. For degree matrix, for The inverse square root matrix is ​​defined as:

[0038] ;

[0039] Where diag represents the diagonal matrix function. Indicates the first Each sample in modality The degree value below, that is, the first The sum of all connection weights of each sample in the adjacency graph;

[0040] Step S25: Utilize graph filtering to fuse features from the user and its neighbors, and enhance neighbor features. The definition is as follows:

[0041] (9),

[0042] in For activation function, The transformation matrix is ​​a learnable matrix in the network. The feature dimensions are those obtained after graph convolution and linear transformation.

[0043] Step S30 includes the following steps:

[0044] Step S31: Different modal characteristics of each AD patient Constructing cross-modal associations to capture complementarity and dependency between different modalities, targeting the first Modal features Generate query vectors through single-layer linear projection. Regarding the first Modal features Generate key vectors through single-layer linear projection. Sum value vector :

[0045] (10)

[0046] in A lightweight projection matrix shared across modes; then the modes are calculated. The correlation score with other modalities is calculated, and the intermodal attention weights are obtained through Softmax normalization, as shown in the following formula:

[0047] (11),

[0048] (12)

[0049] in Indicates the patient Modal and The degree of correlation between modes satisfies Softmax is the activation function. For feature embedding dimension;

[0050] Step S32: For each mode The enhanced features are obtained by aggregating all modal information based on the cross-attention weights. :

[0051] (13);

[0052] Step S33: For each patient The attention enhancement features from all modalities are concatenated to generate the input vector for the gating network. :

[0053] (14)

[0054] Where Concat is a concatenation operation along the feature dimension;

[0055] Next, Input the data into a gated network, perform classification using the softmax activation function, and dynamically generate the gate weights. :

[0056] (15)

[0057] in ,satisfy , Indicates sample In modality The gating weights are as follows; It is a multilayer perceptron neural network;

[0058] Finally, based on the gating mechanism, the strength of the discriminative power of different modalities of data for Alzheimer's disease pathological staging in each sample is adaptively determined to achieve multimodal feature fusion and enhance the feature matrix. , of which Multimodal fusion feature vector of each sample The definition is as follows:

[0059] (16).

[0060] Step S40 includes the following steps:

[0061] Step S41: Collect the adjacency matrix set generated in step S23 for each mode. Concatenate the adjacency matrices of each modality along the modal dimension to construct a three-dimensional adjacency tensor. :

[0062] (17)

[0063] Concat represents the concatenation operation. Representing a three-dimensional adjacency tensor In the middle, modality Lower sample With sample Adjacency weights between them Indicates the first Adjacency matrix of each mode The Middle Line number The elements of the column are equivalent in meaning;

[0064] Step S42: Based on the modal sample-level weights output by the gating network in step S33... Adjacent tensors The initial global adjacency matrix is ​​generated by weighting and fusing the cross-modal attention gating weights. :

[0065] (18)

[0066] in Representation matrix The element in the i-th row and j-th column;

[0067] Step S43: Design cross-modal conflict coefficients Measuring patient response The degree of adjacency conflict between different modes, and the conflict matrix The formula is constructed based on the joint trust and adjacency dissimilarity of patient samples across different modalities, as follows:

[0068] (19)

[0069] in This represents the element in the i-th row and j-th column of matrix K;

[0070] Step S44: Correct the initial adjacency matrix using the conflict coefficient, suppress the connection weights of high-conflict sample pairs, and strengthen the associations of low-conflict sample pairs, so that the global adjacency matrix more accurately reflects the comprehensive pathological similarity of Alzheimer's patients, thus obtaining the global adjacency matrix. :

[0071] (20);

[0072] Where γ is a hyperparameter that controls the decay rate;

[0073] Step S45, based on The value range distribution, referencing three-way decision theory, is used for each sample. The neighbor set is divided differently; a preset threshold is used. ,in The lower bound of the positive domain. The upper limit of the negative threshold is based on the sample. Global similarity Sample The neighbors are divided into positive domains Boundary domain and negative domain ,as follows:

[0074] (twenty one);

[0075] Step S46: Adopt a differentiated comparative learning strategy for different regions: The inner neighbor adopts a hard positive contrast constraint; for The inner neighbor constraint is a confidence-weighted soft positive contrast constraint, based on... The weights are flexibly brought closer together. Intrasamples and samples Forming negative sample pairs; samples Confidence level stratified contrast loss The definition is as follows:

[0076] (twenty two),

[0077] in For cosine similarity calculation, Temperature coefficient, weighting coefficient The definition is as follows:

[0078] (twenty three),

[0079] Facing samples without neighbors That is, only Regions, enhanced samples constructed by adding noise Instance-level comparative learning is employed, with samples... Instance-level contrast loss The definition is as follows:

[0080] (twenty four),

[0081] (25)

[0082] in, Indicates sample The noise enhancement characteristics, Represents the noise enhancement function. This is the noise intensity hyperparameter.

[0083] In step S46, compare the loss. Defined as:

[0084] (26)

[0085] in Represents the set of samples that have neighbors. , Represents the set of samples without neighbors. .

[0086] Step S50 includes the following steps:

[0087] Step S51: Minimize the joint contrast loss With reconstruction loss Iteratively update model parameters to simultaneously achieve cross-modal semantic alignment and feature fidelity, with the total model loss... The definition is as follows:

[0088] (27)

[0089] in For hyperparameters;

[0090] Step S52: Based on the unified representation extracted by the optimized model, the k-means clustering algorithm is used to stage Alzheimer's disease and minimize the clustering loss. The formula is as follows:

[0091] (28)

[0092] in For the number of period categories, Indicates the first The sample set of each cluster For clusters The center of mass, For the sample With center of mass The Euclidean square distance.

[0093] The present invention also provides an electronic device, including a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method.

[0094] The present invention also provides a storage medium storing a computer program or instructions that, when the computer program or instructions are run on a computer, execute the steps of the method described.

[0095] This method eliminates the reliance on a single data source and enhances the depth of fused representations, effectively improving the model's ability to distinguish cases at the staging boundary. This helps doctors overcome the experience limitations of traditional staging and improve the objectivity and reliability of staging judgments.

[0096] Beneficial effects: (1) The neighbor feature fusion method provided by this invention innovatively transforms the problem of fixed neighbor parameter setting into an adaptive fuzzy particle sphere construction problem, that is, mapping samples from the original feature space to a data-driven semantic space that can reveal nonlinear local manifolds. In this space, a pathological adjacency graph is constructed by calculating the fuzzy co-occurrence intensity of samples to multi-scale particles, so that the definition of connection weight changes from the traditional "numerical proximity" to "structural co-occurrence". This not only realizes dynamic adaptation to data density differences, but also effectively suppresses noise interference by utilizing the spatial smoothing effect of particles, thereby significantly enhancing the robustness and structural representation ability of neighbor relationships in complex pathological data scenarios.

[0097] (2) The cross-attention gating network is used to calculate the sample-level view weights, transforming the traditional mode that only considers the semantic information of a single view into a modeling mechanism for interaction between views. By capturing the correlation and complementarity between multimodal features, the problem of neglecting modal association in view weight allocation is solved, and the discrimination ability of multi-view feature fusion is improved.

[0098] (3) Based on the three-branch decision theory, a three-domain differential contrastive learning strategy for cross-modal conflict perception is designed. The unified neighborhood processing paradigm is transformed into a refined partitioning mechanism of positive domain, boundary domain and negative domain. The global adjacency matrix is ​​modified by the conflict coefficient to suppress modal conflict interference. For different decision domains, a differential learning strategy of hard positive constraint, confidence-weighted soft positive constraint and negative sample pair is adopted. At the same time, noise-enhanced instance-level contrast is introduced to solve the problem of learning samples without neighbors, and the reinforcement learning of the discriminative features of boundary samples is realized. Attached Figure Description

[0099] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0100] Figure 1 This is an overall flowchart of the method of the present invention.

[0101] Figure 2 This is a diagram illustrating the overall data processing framework of the method of the present invention.

[0102] Figure 3 This is a flowchart of the method of the present invention.

[0103] Figure 4 This invention provides a framework diagram for a multimodal data-based accurate staging system for Alzheimer's disease. Detailed Implementation

[0104] like Figure 1 , Figure 2 , Figure 3 As shown, this embodiment of the invention provides a multimodal AD staging three-branch contrastive clustering method guided by multi-sphere co-occurrence, including the following steps:

[0105] S10. Acquire multimodal data of AD patients and construct independent encoder-decoder module groups. Drive model learning by minimizing reconstruction loss to ensure that low-dimensional features retain key information of each modality.

[0106] S20. Based on the particle-sphere computation theory, a fuzzy particle-sphere neighborhood is constructed for each modality data. The local structural correlation of samples is captured by the multi-particle-sphere co-occurrence similarity measure, thereby achieving intra-modal feature enhancement and noise suppression.

[0107] S30. Utilize the cross-attention mechanism to model complementary correlations between modalities, and dynamically allocate modal contribution weights through adaptive gating units to achieve deep interaction and sample-level fusion of heterogeneous features;

[0108] S40. Based on the conflict correction fusion similarity matrix, the sample space is divided by three-way decision, and a confidence hierarchical comparison learning strategy is designed for different decision domains to strengthen the discriminative feature learning of boundary samples.

[0109] S50, Joint Contrast Loss With reconstruction loss The model is iteratively optimized, and the final output is the staged clustering results of Alzheimer's patients.

[0110] Step S10 includes the following steps:

[0111] S11. Obtain a multimodal dataset of AD patients. ,in This represents the total number of AD patient samples. Indicates the first A sample of AD patients, where i takes values ​​from 1 to n;

[0112] The i-th AD patient sample Represented as a multidimensional feature vector ,in Representative sample Structural feature vectors in structural magnetic resonance imaging Representative sample positron emission tomography metabolic feature vector of fluorodeoxyglucose. Representative sample The positron emission tomography (PET) characteristic vector of β-amyloid protein characterizes the Aβ deposition load in brain regions. Representative sample cerebrospinal fluid / blood biomarker vectors, among which, and The concentration values ​​of β-amyloid protein with 42 residues and 40 residues, respectively; This represents the total tau protein concentration. This represents the concentration of tau protein phosphorylated at threonine 217. Representative sample The apolipoprotein E genotype vector, representing the number of ε4 alleles, Representative sample The cognitive function scoring vector, with each component corresponding to a quantitative score from the Mini-Mental State Examination (MMSE), Montreal Cognitive Assessment (MoCA), the Alzheimer's Disease Rating Scale-Sum Box (CDR-SB), and the Alzheimer's Disease Assessment Scale-Cognitive Subscale (ADAS-Cog). It also satisfies the original feature total dimension. ,in, Let represent the real number space, where the superscript indicates the dimension of the corresponding vector, and T denotes the transpose of the vector. These are the characteristic dimensions of sMRI, FDG-PET, and Aβ-PET, respectively; 3, 1, and 4 are the fixed dimensions of biomarkers, genotypes, and cognitive scores.

[0113] In this embodiment, the original feature dimensions of each modality are processed according to the standard workflow of the Alzheimer's Disease Neuroimaging Initiative (ADNI) database, using the AAL-90 brain region segmentation template. .

[0114] S12. Construct independent coding modules for each modality. and decoder module ,in Indicates the total number of modalities and each encoding module. Mapping the original features to a latent representation of uniform dimension; each decoder module The latent representation is reconstructed into the original feature space, where For network parameters, For modality The original feature dimension, For the unified dimensions;

[0115] The encoding process generates a low-dimensional feature set through reparameterization techniques. The definition is as follows:

[0116] (1),

[0117] in , To obtain from the multivariate standard normal distribution The noise vector sampled in the middle, This represents a normal distribution with a mean of zero and a covariance equal to the identity matrix. These are the mean and variance vectors of the output, respectively. For the sample In the Semantic vectors in each modality This is an element-wise multiplication operation;

[0118] The latent features after unifying each modality are shown in Table 1 below.

[0119] Table 1

[0120]

[0121] In this embodiment, there are a total of 8 samples with a unified feature dimension. For the sake of simplicity, only some representative metrics for each modality are listed; the actual encoder output is a complete feature vector.

[0122] The decoding process will represent the latent representation Reconstructing Back to the First The original feature space of each modality :

[0123] (2),

[0124] The reconstruction loss calculation uses the mean squared error to measure the difference between the original features and the reconstructed features, and minimizes the reconstruction loss. Optimize encoder parameters and decoder parameters , making the potential representation Able to preserve original modal features to the greatest extent The semantic information is expressed in the following formula:

[0125] (3).

[0126] For ease of understanding, the sample is described here. The reconstruction loss in Mode 1 will be described in detail, including the original features. Reconstructing features The reconstruction loss is The reconstruction loss of the entire dataset Reconstruct the loss for all samples across all modalities.

[0127] Step S20 includes the following steps:

[0128] S21, in each mode low-dimensional feature subspace In the middle, order Indicates based on The generated set of spheres, in which Indicates the total number of balls. , For granules Number of samples center and radius The definition is as follows:

[0129] (4),

[0130] First, initialize the entire dataset as a single sphere, if the partitioning condition is met. and Then, 2-means is used to split the granule into two sub-granules. and Distribution measure The definition is as follows:

[0131] (5);

[0132] In mode 1, after the above iterative splitting, four spheres are generated. ,in .

[0133] S22. For each sample in the m-th modality Calculate the membership degree of the sample to the top-p nearest spheres, and set the membership degree of the sample to 0 for the remaining spheres. In mode m, the j-th sample... For the i-th ball membership degree The definition is as follows:

[0134] (6),

[0135] in It is a hyperparameter that adjusts the degree of blur. It is a sample of Nearest neighbor set of particles, membership matrix exp represents the natural exponential function;

[0136] In this embodiment, take The membership matrix for mode 1 can be calculated as follows:

[0137] ;

[0138] S23. Calculate the fuzzy co-occurrence similarity between samples to construct a pathological adjacency graph: for any two samples and ,if and For the same ball membership degree and If all values ​​are not less than the preset threshold η, it indicates that the value is within the range of... Patients are defined as being at similar stages of disease progression within a localized pathological space at a given scale. With patients Connection weights between For the fuzzy Jaccard similarity on all spheres:

[0139] (7),

[0140] In this example, we take η = 0.7. The adjacency matrix of mode 1 is calculated as follows:

[0141] ;

[0142] S24, to Normalization is performed to obtain the normalized Laplace matrix. :

[0143] (8),

[0144] in It is the identity matrix. For degree matrix, for The inverse square root matrix is ​​defined as:

[0145] ;

[0146] Where diag represents the diagonal matrix function. Indicates the first Each sample in modality The degree value is the sum of all connection weights of the sample in the adjacency graph;

[0147] In this embodiment, the mode 1 normalized Laplace matrix can be calculated. :

[0148] ;

[0149] S25. Utilize graph filtering to fuse features from the user and its neighbors, thereby enhancing neighbor features. The definition is as follows:

[0150] (9),

[0151] in For activation function, The transformation matrix is ​​a learnable matrix in the network. The feature dimensions are those obtained after graph convolution and linear transformation.

[0152] Step S30 includes the following steps:

[0153] S31. Different modal characteristics for each AD patient Constructing cross-modal associations to capture complementarity and dependency between different modalities, targeting the first Modal features Generate query vectors through single-layer linear projection. Regarding the first Modal features Generate key vectors through single-layer linear projection. Sum value vector :

[0154] (10)

[0155] in A lightweight projection matrix shared across modes; then the modes are calculated. The correlation score with other modalities is calculated, and the intermodal attention weights are obtained through Softmax normalization, as shown in the following formula:

[0156] (11),

[0157] (12)

[0158] in Indicates the patient Modal and The degree of correlation between modes satisfies Softmax is the activation function. For feature embedding dimension;

[0159] S32, For each mode The enhanced features are obtained by aggregating all modal information based on the cross-attention weights. :

[0160] (13);

[0161] S33, For each patient The attention enhancement features from all modalities are concatenated to generate the input vector for the gating network. :

[0162] (14)

[0163] Where Concat is a concatenation operation along the feature dimension;

[0164] Next, Input the data into a gated network, perform classification using the softmax activation function, and dynamically generate the gate weights. :

[0165] (15)

[0166] in ,satisfy , Indicates sample In modality The gating weights are as follows; For a multilayer perceptron neural network, given the enhanced features of sample 1 as input, the gating weights for sample 1 are calculated as follows: .

[0167] Finally, based on the gating mechanism, the strength of the discriminative power of different modalities of data for Alzheimer's disease pathological staging in each sample is adaptively determined to achieve multimodal feature fusion and enhance the feature matrix. , of which Multimodal fusion feature vector of each sample The definition is as follows:

[0168] (16)

[0169] Then the enhanced features of sample 1 are: .

[0170] Step S40 includes the following steps:

[0171] S41. Collect the adjacency matrix set generated in each modal step S23. Concatenate the adjacency matrices of each modality along the modal dimension to construct a three-dimensional adjacency tensor. :

[0172] (17)

[0173] in Representing a three-dimensional adjacency tensor In the middle, modality Lower sample With sample Adjacency weights between them Indicates the first Adjacency matrix of each mode The Middle Line number The elements of the column are equivalent in meaning;

[0174] S42. Based on the modal sample-level weights output by the gating network in step S33. Adjacent tensors The initial global adjacency matrix is ​​generated by weighting and fusing the cross-modal attention gating weights. :

[0175] (18)

[0176] in Representation matrix The element in the i-th row and j-th column;

[0177] S43, Design cross-modal conflict coefficient Measuring patient response The degree of adjacency conflict between different modes, and the conflict matrix The formula is constructed based on the joint trust and adjacency dissimilarity of patient samples across different modalities, as follows:

[0178] (19)

[0179] in This represents the element in the i-th row and j-th column of matrix K;

[0180] S44. The initial adjacency matrix is ​​corrected using the conflict coefficient to suppress the connection weights of high-conflict sample pairs and strengthen the associations of low-conflict sample pairs, so that the global adjacency matrix more accurately reflects the comprehensive pathological similarity of Alzheimer's patients, thus obtaining the global adjacency matrix. In this example, γ=1 is taken, and the calculation yields:

[0181] (20);

[0182] S45, based on The value range distribution, referencing three-way decision theory, is used for each sample. The neighbor set is divided differently; a preset threshold is used. ,in The lower bound of the positive domain. The upper limit of the negative threshold is based on the sample. Global similarity Sample The neighbors are divided into positive domains Boundary domain and negative domain The formula is as follows:

[0183] (twenty one);

[0184] In this embodiment For sample 1,

[0185] S46. Adopt differentiated comparative learning strategies for different regions: The inner neighbor adopts a hard positive contrast constraint; for The inner neighbor constraint is a confidence-weighted soft positive contrast constraint, based on... The weights are flexibly brought closer together. Intrasamples and samples Forming negative sample pairs; samples Confidence level stratified contrast loss The definition is as follows:

[0186] (twenty two),

[0187] in For cosine similarity calculation, Temperature coefficient, weighting coefficient The definition is as follows:

[0188] (twenty three),

[0189] Facing samples without neighbors That is, only Regions, enhanced samples constructed by adding noise Instance-level comparative learning is employed, with samples... Instance-level contrast loss The definition is as follows:

[0190] (twenty four),

[0191] (25)

[0192] in, Indicates sample The noise enhancement characteristics, Represents the noise enhancement function. This refers to the noise intensity hyperparameter.

[0193] Comparative loss Defined as:

[0194] (26)

[0195] in Represents the set of samples that have neighbors. , Represents the set of samples without neighbors. .

[0196] Step S50 includes the following steps:

[0197] S51, By minimizing the joint contrast loss With reconstruction loss Iteratively update model parameters to simultaneously achieve cross-modal semantic alignment and feature fidelity, with the total model loss... The definition is as follows:

[0198] (27)

[0199] In this embodiment =1 is a hyperparameter;

[0200] S52. Based on the unified representation extracted by the optimized model, the k-means clustering algorithm is used to stage Alzheimer's disease and minimize the clustering loss. The formula is as follows:

[0201] (28)

[0202] In this embodiment, It is divided into four stages: CN (normal cognition), EMCI (early mild cognitive impairment), LMCI (late-stage mild cognitive impairment), and AD. Indicates the first The sample set of each cluster For clusters The center of mass, For the sample With center of mass The Euclidean square distance.

[0203] Example 2: In this example, all hyperparameters involved in each process (including but not limited to the number of nearest neighbor particles, upper and lower limits of the boundary domain, and other control parameters related to modal characteristics) are set as preset initial values. These hyperparameters can be adaptively and iteratively optimized by considering factors such as sample size, data modal characteristics, and indicators like loss function convergence and staging accuracy during model training. Through periodic evaluation and dynamic update mechanisms, the values ​​of each hyperparameter are adjusted to continuously improve staging accuracy and model stability while maintaining computational complexity.

[0204] Example 3: This example provides a system for accurate staging of multimodal Alzheimer's disease data. For example...Figure 4 As shown, it includes modules for data acquisition, intra-modal feature enhancement, inter-modal feature enhancement, sample neighbor partitioning, contrastive learning, clustering, and result output.

[0205] This invention provides a multimodal AD staging three-branch contrastive clustering method guided by multi-sphere co-occurrence. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A multimodal AD staging three-branch contrastive clustering method guided by multi-sphere co-occurrence, characterized in that, Includes the following steps: Step S10: Obtain multimodal clinical data of AD patients and construct independent encoding and decoding module groups, and drive model learning by minimizing reconstruction loss; Step S20: Construct fuzzy particle sphere neighborhoods for each modality based on particle sphere computation theory, capture local structural associations of samples through multi-particle sphere co-occurrence similarity measurement, and achieve intra-modal feature enhancement and noise suppression; Step S30: Construct a cross-modal gated attention fusion network, model the complementary correlation between modalities using the cross-attention mechanism, and dynamically allocate modal contribution weights through adaptive gating units to complete the deep interaction and sample-level fusion of heterogeneous features; Step S40: Based on the conflict correction fusion similarity matrix, the sample space is divided into positive domain, negative domain and boundary domain through three-way decision. A confidence hierarchical comparison learning strategy is designed for different decision domains to strengthen the discriminative feature learning of boundary samples. Step S50, Joint comparison loss With reconstruction loss The model is iteratively optimized, and the final output is the staged clustering results of Alzheimer's patients.

2. The method according to claim 1, characterized in that, Step S10 includes the following steps: Step S11: Obtain multimodal data of AD patients, including structural magnetic resonance imaging and fluorodeoxyglucose positron emission tomography neuroimaging data, molecular biomarkers, and cognitive assessment scores of the Mini-Mental State Examination and Alzheimer's Disease Rating Scale. Multimodal datasets ,in This represents the total number of AD patient samples. Indicates the first A sample of AD patients, where i takes values ​​from 1 to n; The i-th AD patient sample Represented as a multidimensional feature vector ,in Representative sample Structural feature vectors of sMRI Representative sample Metabolic feature vectors of fluorinated deoxyglucose positron emission tomography (FDG-PET). Representative sample The positron emission tomography (PET) feature vector of β-amyloid protein (Aβ-PET) characterizes the Aβ deposition load in brain regions. Representative sample cerebrospinal fluid / blood biomarker vectors, among which, and The concentration values ​​of β-amyloid protein with 42 residues and 40 residues, respectively; This represents the total tau protein concentration. This represents the concentration of tau protein phosphorylated at threonine 217. Representative sample apolipoprotein E genotype vector, Representative sample Cognitive function score vector, each component , The quantitative scores correspond sequentially to the Brief Mental Status Examination, the Montreal Cognitive Assessment, the Alzheimer's Disease Rating Scale (AMR) summary frame, and the Alzheimer's Disease Rating Scale - Cognitive Subscale, and satisfy the total dimensions of the original features. , Let T denote the real number space, and let T denote the vector transpose. These are the characteristic dimensions of sMRI, FDG-PET, and Aβ-PET, respectively. Step S12: Construct independent coding modules for each modality and decoder module ,in Indicates the total number of modalities and each encoding module. Mapping the original features to a latent representation of uniform dimension; each decoder module The latent representation is reconstructed into the original feature space, where For network parameters, Let m be the original feature dimension. For the unified dimensions; The encoding process generates a low-dimensional feature set through reparameterization techniques. The definition is as follows: （1）， in , To obtain from the multivariate standard normal distribution The noise vector sampled in the middle, This represents a normal distribution with a mean of zero and a covariance equal to the identity matrix. These are the mean and variance vectors of the output, respectively. For the sample In the Semantic vectors in each modality This is an element-wise multiplication operation; The decoding process will represent the latent representation Reconstructing Back to the First The original feature space of each modality : （2）， The reconstruction loss calculation uses the mean squared error to measure the difference between the original features and the reconstructed features, and minimizes the reconstruction loss. Optimize encoder parameters and decoder parameters , making the potential representation It can preserve the original modal features to the greatest extent. Semantic information.

3. The method according to claim 2, characterized in that, In step S12, the reconstruction loss The formula is as follows: （3）。 4. The method according to claim 3, characterized in that, Step S20 includes the following steps: Step S21, in each mode low-dimensional feature subspace In the middle, order Indicates based on The generated set of spheres, in which Indicates the total number of balls. , For granules Number of samples center and radius The definition is as follows: （4）， First, initialize the entire dataset as a single sphere, if the partitioning condition is met. and Then, 2-means is used to split the granule into two sub-granules. and Distribution measure The definition is as follows: （5）； Step S22: For each sample in the m-th mode ,calculate The membership degree of the top-p closest spheres in terms of geometric distance, and the sample The membership degree of the remaining spheres is set to 0; in mode m, the j-th sample For the i-th ball membership degree The definition is as follows: （6）， in It is a hyperparameter that adjusts the degree of blur. It is a sample of Nearest neighbor set of particles, membership matrix ;exp represents the natural exponential function; Step S23: Calculate the fuzzy co-occurrence similarity between samples to construct a pathological adjacency graph: For any two samples and ,if and For the same ball membership degree and If all values ​​are not less than the preset threshold η, it indicates that the value is within the range of... Patients are defined as being at similar stages of disease progression within a localized pathological space at a given scale. With patients Connection weights between For the fuzzy Jaccard similarity on all spheres: （7）， in, It is a smoothing term, graph adjacency matrix ; Step S24, for Normalization is performed to obtain the normalized Laplace matrix. : （8）， in It is the identity matrix. For degree matrix, for The inverse square root matrix is ​​defined as: ； Where diag represents the diagonal matrix function. Indicates the first Each sample in modality The degree value below; Step S25, Neighbor Enhancement Features The definition is as follows: （9）， in For activation function, The transformation matrix is ​​a learnable matrix in the network. The feature dimensions are those obtained after graph convolution and linear transformation.

5. The method according to claim 4, characterized in that, Step S30 includes the following steps: Step S31: Different modal characteristics of each AD patient Constructing cross-modal associations to capture complementarity and dependency between different modalities, targeting the first Modal features Generate query vectors through single-layer linear projection. Regarding the first Modal features Generate key vectors through single-layer linear projection. Sum value vector : （10）， in A lightweight projection matrix shared across modes; then the modes are calculated. The correlation score with other modalities is calculated, and the intermodal attention weights are obtained through Softmax normalization, as shown in the following formula: （11）， （12）， in Indicates the patient Modal and The degree of correlation between modes satisfies Softmax is the activation function. For feature embedding dimension; Step S32: For each mode The enhanced features are obtained by aggregating all modal information based on the cross-attention weights. : （13）； Step S33: For each patient The attention enhancement features from all modalities are concatenated to generate the input vector for the gating network. : （14）， Where Concat is a concatenation operation along the feature dimension; Next, Input the data into a gated network, perform classification using the softmax activation function, and dynamically generate the gate weights. : （15）， in ,satisfy , Indicates sample In modality The gating weights are as follows; It is a multilayer perceptron neural network; Finally, based on the gating mechanism, the strength of the discriminative power of different modalities of data for Alzheimer's disease pathological staging in each sample is adaptively determined to achieve multimodal feature fusion and enhance the feature matrix. , of which Multimodal fusion feature vector of each sample The definition is as follows: （16）。 6. The method according to claim 5, characterized in that, Step S40 includes the following steps: Step S41: Collect the adjacency matrix set generated in step S23 for each mode. Concatenate the adjacency matrices of each modality along the modal dimension to construct a three-dimensional adjacency tensor. : （17）， Concat represents the concatenation operation. Representing a three-dimensional adjacency tensor In the middle, modality Lower sample With sample Adjacency weights between them Indicates the first Adjacency matrix of each mode The Middle Line 1 Column elements; Step S42: Based on the modal sample-level weights output by the gating network in step S33... Adjacent tensors The initial global adjacency matrix is ​​generated by weighting and fusing the cross-modal attention gating weights. : （18）， in Representation matrix The element in the i-th row and j-th column; Step S43: Design cross-modal conflict coefficients Measuring patient response The degree of adjacency conflict between different modes, and the conflict matrix The formula is constructed based on the joint trust and adjacency dissimilarity of patient samples across different modalities, as follows: （19）， in This represents the element in the i-th row and j-th column of matrix K; Step S44: Correct the initial adjacency matrix using the conflict coefficient, suppress the connection weights of high-conflict sample pairs, and strengthen the associations of low-conflict sample pairs, so that the global adjacency matrix more accurately reflects the comprehensive pathological similarity of Alzheimer's patients, thus obtaining the global adjacency matrix. : （20）； Where γ is a hyperparameter; Step S45, based on The value range distribution, referencing three-way decision theory, is used for each sample. The neighbor set is divided differently; a preset threshold is used. ,in The lower bound of the positive domain. The upper limit of the negative threshold is based on the sample. Global similarity Sample The neighbors are divided into positive domains Boundary domain and negative domain ,as follows: （21）； Step S46: Adopt a differentiated comparative learning strategy for different regions: The inner neighbor adopts a hard positive contrast constraint; for The inner neighbor constraint is a confidence-weighted soft positive contrast constraint, based on... The weights are flexibly brought closer together. Intrasamples and samples Forming negative sample pairs; samples Confidence level stratified contrast loss The definition is as follows: （22）， in For cosine similarity calculation, Temperature coefficient, weighting coefficient The definition is as follows: （23）， Facing samples without neighbors That is, only Regions, enhanced samples constructed by adding noise Instance-level contrastive learning is employed, with samples... Instance-level contrast loss The definition is as follows: （24）， （25）， in, Indicates sample The noise enhancement characteristics, Represents the noise enhancement function. This is the noise intensity hyperparameter.

7. The method according to claim 6, characterized in that, In step S46, compare the loss. Defined as: （26）， in Represents the set of samples that have neighbors. , Represents the set of samples without neighbors. .

8. The method according to claim 7, characterized in that, Step S50 includes the following steps: Step S51: Minimize the joint contrast loss With reconstruction loss Iteratively update model parameters to simultaneously achieve cross-modal semantic alignment and feature fidelity, with the total model loss... The definition is as follows: （27）， in For hyperparameters; Step S52: Based on the unified representation extracted by the optimized model, the k-means clustering algorithm is used to stage Alzheimer's disease and minimize the clustering loss. The formula is as follows: （28）， in For the number of period categories, Indicates the first The sample set of each cluster For clusters The center of mass, For the sample With center of mass The Euclidean square distance.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 8.

10. A storage medium, characterized in that, It stores a computer program or instructions that, when run on a computer, perform the steps of the method as described in any one of claims 1 to 8.

Images