A mild cognitive impairment classification method and system based on dynamic heterogeneous graph fusion

Abstract

This invention relates to the field of medical image processing technology, proposing a classification method and system for mild cognitive impairment based on dynamic heterogeneous graph fusion. The method includes: preprocessing acquired data to obtain standardized genetic features and standardized image features; grouping and reducing the dimensionality of the standardized genetic features to output low-dimensional genetic features; outputting complete image features based on the low-dimensional genetic features and standardized image features through a bidirectional generative adversarial network; constructing a heterogeneous graph from the low-dimensional genetic features and complete image features, the heterogeneous graph containing four edge types and retaining strongly correlated edges; using a two-layer heterogeneous graph transformer for convolutional extraction of node features, inputting these features into a global average pooling layer to obtain gene modality embeddings and brain region modality embeddings, inputting these features into a perceptron to learn dynamic fusion weights, and outputting global fusion features; classifying the global fusion features and outputting the classification result for mild cognitive impairment. This invention solves the problem of existing classification methods easily losing key classification features.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a classification method and system for mild cognitive impairment based on dynamic heterogeneous image fusion. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Alzheimer's disease is a neurodegenerative disease prevalent among middle-aged and elderly people. As its primary precursor stage, accurate classification of mild cognitive impairment (MCI) is a core prerequisite for early intervention. The accurate differentiation between normal cognitive individuals (NC), early mild cognitive impairment (EMCI), and late mild cognitive impairment (LMCI) directly determines the formulation of clinical intervention plans and the assessment of disease progression risk. Advances in genetic imaging technology have made the fusion analysis of single nucleotide polymorphism (SNP) genetic data and functional magnetic resonance imaging (fMRI) data an important technological direction for improving the accuracy of MCI classification. Combining these two methods can capture disease-related characteristics from multiple dimensions, providing more comprehensive data support for multi-classification tasks of MCI.

[0004] Existing genetic imaging-based MCI classification methods still suffer from technical defects that affect classification accuracy and robustness: existing methods are difficult to accurately model the multi-level heterogeneous associations between genes and image data, and mostly adopt fixed-weight fusion strategies, which cannot adaptively achieve the optimal integration of the two types of features, easily causing the loss or dilution of key classification features, thus limiting further improvement in classification accuracy. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, the present invention provides a classification method and system for mild cognitive impairment based on dynamic heterogeneous graph fusion, which solves the problem that existing classification methods are prone to losing key classification features.

[0006] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of this invention provides a method for classifying mild cognitive impairment based on dynamic heterogeneous graph fusion, comprising: Preprocessing of single nucleotide polymorphism gene data and functional magnetic resonance imaging data yields standardized gene features and standardized image features; Standardized gene features are grouped and dimensionality reduced to preserve the topological structure of gene data and output low-dimensional genetic features. Based on the aforementioned low-dimensional genetic features and standardized image features, a bidirectional generative adversarial network is used to complete the missing image features and output the completed image features. The low-dimensional genetic features and the completed image features are used to construct a heterogeneous graph. The heterogeneous graph includes gene nodes and brain region nodes, as well as four edge types: gene-gene, brain region-brain region, gene-brain region, and brain region-gene, retaining strongly correlated edges. A two-layer heterogeneous graph transformer is used to extract node features through convolution. These features are then input into a global average pooling layer to obtain gene modality embeddings and brain region modality embeddings. The perceptron is then input to learn dynamic fusion weights, and the global fusion features are output. The global fusion features are classified to output a classification result for mild cognitive impairment.

[0007] As a further technical solution, the preprocessing step includes: Single nucleotide polymorphism (SNP) gene data preprocessing: PLINK tool was used for sample quality control, sex verification, minimum allele frequency filtering, and Hardy-Weinberg balance filtering. The data were grouped according to biological attributes, and target genes and corresponding SNP sites were selected and encoded as 0, 1, and 2. Functional magnetic resonance imaging (fMRI) image data preprocessing employed the DPABI toolkit for temporal correction, head motion correction, spatial normalization, and smoothing. Time-series features of each brain region were extracted based on brain region templates.

[0008] As a further technical solution, the grouping dimensionality reduction step includes: dividing the gene features according to biological attributes, using the K-nearest neighbor local linear embedding algorithm to reduce the dimensionality of each group, and splicing the dimensionality reduction results of each group to generate low-dimensional genetic features.

[0009] As a further technical solution, the bidirectional generative adversarial network includes a first generator, a second generator, a first discriminator, and a second discriminator; the first generator generates completed image features based on low-dimensional genetic features, and the second generator generates verification genetic features based on real image features; the first discriminator identifies the authenticity of the completed image features, and the second discriminator identifies the authenticity of the verification genetic features; the parameters of the first generator and the second generator are updated synchronously by calculating adversarial loss, cross-modal topological consistency loss, and reconstruction loss.

[0010] As a further technical solution, the two-layer heterogeneous graph transformer convolution calculates the attention weights between gene nodes and brain region nodes through a type-aware attention mechanism, and weights and fuses the features of adjacent nodes through a feature aggregation function to output the hidden features of the nodes.

[0011] As a further technical solution, the extracted node features are obtained by global average pooling to obtain gene modality embeddings and brain region modality embeddings. The concatenated gene modality embeddings and brain region modality embeddings are then passed through a first fully connected layer, a ReLU activation layer, and a second fully connected layer in sequence to output gene modality weight scores and brain region modality weight scores. The dynamic fusion weights are obtained by Softmax normalization. The gene modality embeddings and brain region modality embeddings are weighted and summed according to the dynamic fusion weights to output global fusion features.

[0012] As a further technical solution, a linear classifier is used to map the global fusion features to a three-dimensional classification score, which is then normalized to a classification probability distribution by a Softmax activation layer, outputting the classification results for normal controls, early mild cognitive impairment, and late mild cognitive impairment.

[0013] A second aspect of the present invention provides a classification system for mild cognitive impairment based on dynamic heterogeneous graph fusion, comprising: The preprocessing module is configured to preprocess single nucleotide polymorphism gene data and functional magnetic resonance imaging data to obtain standardized gene features and standardized image features. The topology-preserving dimensionality reduction module is configured to: group and reduce the dimensionality of standardized gene features, preserve the topological structure of the gene data, and output low-dimensional genetic features; The cross-modal generative completion module is configured to: based on the low-dimensional genetic features and standardized image features, complete the missing image features through a bidirectional generative adversarial network and output the completed image features; The dynamic heterogeneous graph fusion module is configured to: construct a heterogeneous graph from the low-dimensional genetic features and the completed image features, wherein the heterogeneous graph includes gene nodes and brain region nodes, and four edge types: gene-gene, brain region-brain region, gene-brain region, and brain region-gene, retaining strongly correlated edges; extract node features using a two-layer heterogeneous graph transformer convolution, input the features into a global average pooling layer to obtain gene modality embeddings and brain region modality embeddings, input the features into a perceptron to learn dynamic fusion weights, and output global fusion features; The classification module is configured to classify the global fusion features and output a classification result for mild cognitive impairment.

[0014] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps of a method for classifying mild cognitive impairment based on dynamic heterogeneous graph fusion as described in the first aspect of the present invention.

[0015] The fourth aspect of the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in a classification method for mild cognitive impairment based on dynamic heterogeneous graph fusion as described in the first aspect of the present invention.

[0016] The above one or more technical solutions have the following beneficial effects: This invention employs topology-preserving dimensionality reduction to group and reduce the dimensionality of high-dimensional gene features. While significantly reducing feature dimensionality and computational redundancy, it fully preserves the core classification features of the gene data, avoiding the problem of losing key information in traditional dimensionality reduction methods and effectively improving the utilization efficiency of gene features. Cross-modal generation and completion achieves accurate completion of missing image features, solving the core problem of decreased classification performance due to missing image modalities commonly found in clinical data acquisition. This significantly improves the system's adaptability and robustness of classification results in real-world clinical applications. Dynamic heterogeneous graph fusion constructs a heterogeneous graph structure with multiple edge types, accurately capturing the association features between genes and brain regions. Simultaneously, adaptive learning of dynamic fusion weights achieves optimal integration of gene and image features. Compared to traditional fixed-weight fusion methods, this method extracts key classification information from multimodal data more efficiently, significantly improving the accuracy of mild cognitive impairment classification.

[0017] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0018] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0019] Figure 1 This is a flowchart of the method in the first embodiment; Figure 2 This is a schematic diagram of the data preprocessing module in the second embodiment; Figure 3 This is a schematic diagram of the topology-preserving dimensionality reduction module in the second embodiment; Figure 4 This is a schematic diagram of the cross-modal generation and completion module in the second embodiment; Figure 5 This is a schematic diagram of the dynamic heterogeneous graph fusion module in the second embodiment; Figure 6 This is a schematic diagram of the classification module in the second embodiment; Figure 7 This is a schematic diagram of the overall system structure of the second embodiment. Detailed Implementation

[0020] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0022] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0023] Existing genetic imaging-based MCI classification methods have many technical shortcomings: Firstly, in the high-dimensional gene data processing stage, existing dimensionality reduction methods are prone to destroying the inherent correlation structure of the data, or fail to effectively optimize the dimensionality reduction of high-dimensional data, resulting in high feature redundancy and large noise interference, which directly reduces the feature quality of subsequent classification tasks.

[0024] Secondly, in the multimodal feature fusion stage, existing methods are difficult to accurately model the multi-level heterogeneous associations between genes and image data. They often adopt fixed weight fusion strategies, which cannot adaptively achieve the optimal integration of the two types of features. This can easily lead to the loss or dilution of key classification features, thus limiting the further improvement of classification accuracy.

[0025] Third, most existing methods are developed based on the assumption of complete multimodal data, without considering the common problem of missing image modalities in clinical scenarios. They cannot maintain stable classification performance in scenarios with incomplete data and are difficult to adapt to the actual clinical application needs.

[0026] This invention aims to effectively improve the accuracy, stability, and clinical scenario adaptability of mild cognitive impairment classification tasks by standardizing multimodal data, optimizing gene feature dimensionality reduction, completing missing image features, and fusing and classifying multimodal heterogeneous features.

[0027] The specific implementation plan adopted to achieve the above objectives will be described in detail below.

[0028] Example 1 like Figure 1 As shown, this embodiment discloses a classification method for mild cognitive impairment based on dynamic heterogeneous graph fusion, including: Step S1: Preprocess the single nucleotide polymorphism gene data and functional magnetic resonance imaging data to obtain standardized gene features and standardized image features.

[0029] Step S2: Group and reduce the dimensionality of standardized gene features, preserve the topological structure of gene data, and output low-dimensional genetic features.

[0030] Step S3: Based on low-dimensional genetic features and standardized image features, the missing image features are completed using a bidirectional generative adversarial network, and the completed image features are output.

[0031] Step S4: Construct a heterogeneous graph by combining low-dimensional genetic features and completed image features. The heterogeneous graph includes gene nodes and brain region nodes, as well as four edge types: gene-gene, brain region-brain region, gene-brain region, and brain region-gene, retaining strongly correlated edges. Use a two-layer heterogeneous graph transformer convolution to extract node features, input them into a global average pooling layer to obtain gene modality embeddings and brain region modality embeddings, input them into a perceptron to learn dynamic fusion weights, and output global fusion features.

[0032] Step S5: Classify the global fusion features and output the classification results for mild cognitive impairment.

[0033] Specifically, like Figure 2 As shown, in step S1, the preprocessing of SNP (single nucleotide polymorphism) gene data is completed using PLINK software. Sample quality control, gender verification, minimum allele frequency filtering and Hardy-Weinberg balance filtering are performed sequentially. Genes are grouped according to biological attributes, target genes and corresponding SNP sites are selected and encoded as 0, 1 and 2, and finally standardized gene features are output.

[0034] fMRI (functional magnetic resonance imaging) image data preprocessing was performed using the DPABI toolkit, sequentially performing temporal correction, head motion correction, spatial standardization, and smoothing. Based on brain region templates, time-series features of each brain region were extracted, and finally, standardized image features were output.

[0035] By standardizing and normalizing the SNP gene data and fMRI image data through the data preprocessing module, data noise and interference can be effectively removed, and high-quality gene and brain region features can be extracted, laying a stable and reliable data foundation for subsequent feature processing and classification tasks.

[0036] like Figure 3 As shown, step S2 includes gene feature grouping and K-nearest neighbor LLE dimensionality reduction. Specifically: In this embodiment, standardized gene features are divided into 90 groups according to biological attributes. Each group of standardized gene features is independently input into a K-nearest neighbor LLE dimensionality reduction unit (K=30) to reduce the dimensionality of each group of high-dimensional features to a 6-dimensional low-dimensional embedding. Then, the 90 groups of low-dimensional embeddings are spliced ​​and integrated to output a global 540-dimensional low-dimensional genetic feature, which preserves the genetic association topology of the gene data while reducing dimensionality.

[0037] like Figure 4 As shown, step S3 includes a bidirectional generative adversarial network and a parameter update unit. Specifically: Low-dimensional genetic feature input generator (SNP→fMRI), generating complete image features; standardized image feature input generator (fMRI→SNP) to generate and verify genetic characteristics.

[0038] Image feature input discriminator Image authenticity is verified, and genetic characteristics are generated and then input into the discriminator. Genetic authenticity is verified; the results are then used for loss function verification, calculating adversarial loss, cross-modal topological consistency loss, and reconstruction loss, with the following formulas: (1) Adversarial loss: By minimizing the probability that the generated features are identified as “false” by the discriminator, the generated data is forced to approximate the true distribution.

[0039] For image modalities, the loss is defined as:

[0040] For gene modalities, the loss is defined as:

[0041] The total losses from the confrontation are:

[0042] In the formula, 1 and 0 represent vectors with all 1s and all 0s, respectively (the dimensions match the number of samples in the batch). To enable the discriminator to analyze images and genetic features The discriminant output. BCELoss is the binary cross-entropy loss. F Indicates image features, Z Indicates gene characteristics, Indicates the generated image features, This indicates the generated gene characteristics.

[0043] (2) Cross-modal topological consistency loss: By constraining the distance distribution of genes, real images and generated images to be consistent, the topological association between modalities is preserved.

[0044] First, calculate the sample distance matrix:

[0045] In the formula, These represent the feature distances of genes, real images, and generated images, respectively; cdist refers to the calculated distance matrix; after normalization...

[0046] Then, the topological loss is defined as:

[0047] Where MSE is the mean squared error.

[0048] (3) Reconstruction Loss: By forcibly generating features and then reconstructing them through reverse generation, the original features can be reconstructed, thus strengthening the bidirectional consistency of modality mapping. For gene features, the reconstruction loss is:

[0049] For image features, the reconstruction loss is:

[0050] The total reconstruction losses are:

[0051] The training of the cross-modal generative completion module employs an alternating optimization strategy: the discriminator and generator are trained using the Adam optimizer (learning rate 0.0002, momentum parameter...). , The parameters are updated, with the discriminator updated once per iteration and the generator updated twice per iteration to enhance learning performance. The total loss of the generator is...

[0052] The weight of 0.1 is used to reduce the interference of topology and reconstruction loss on adversarial learning.

[0053] Step S4 includes heterogeneous graph construction, heterogeneous graph feature extraction, and dynamic multimodal attention fusion. Specifically: Low-dimensional genetic features and completed image features are input into a dynamic heterogeneous map fusion model, such as... Figure 5 As shown, the model includes a heterogeneous graph construction module, two heterogeneous graph input modules, a global average pooling module, a modality embedding input module, and a global fusion module connected in sequence.

[0054] (1) Heterogeneous graph construction module The heterogeneous graph construction module maps low-dimensional genetic features and completed image features to gene nodes and brain region nodes, respectively, and constructs four types of edges: gene-gene, brain region-brain region, gene-brain region, and brain region-gene. Strongly correlated edges with an absolute value of Pearson correlation coefficient greater than 0.3 are retained to complete the heterogeneous graph construction.

[0055] (2) Heterogeneous graph input module The heterogeneous graph input module consists of a sequentially connected HGTConv layer, a batch normalization layer, a ReLU activation layer, and a Dropout layer. Its purpose is to extract deep features from nodes, obtaining gene modality embeddings and brain region modality embeddings. Specifically: The HGTConv layer (Heterogeneous Graph Transformer Convolutional Layer) is the core layer for heterogeneous graph feature extraction. Unlike traditional graph convolutions that only process nodes / edges of the same type, the HGTConv layer can accurately identify the heterogeneous attribute differences between "gene nodes and brain region nodes" and calculate the attention weights between different types of node pairs through a type-aware attention mechanism. The core of HGTConv is an adaptive attention mechanism based on node and edge types, which focuses on the node features of the m-th layer. Its update formula is:

[0056] Where B is the set of edge types ({gg,rr,gr,rg}). For a set of neighboring nodes of type b connected to p, MLP is a multilayer perceptron, i.e., for features... Perform nonlinear transformations. Type-aware attention weights:

[0057] here and For a specific parameter of type b, || denotes vector concatenation, and LN denotes layer normalization.

[0058] After each HGTConv feature extraction layer, a batch normalization layer is connected. Its core function is to standardize the node features within the batch (adjusting the feature mean to 0 and the variance to 1), avoiding drastic changes in the input distribution of the subsequent layer due to the update of the parameters of the previous layer, greatly improving the training speed of the model, while reducing the sensitivity of the model to the initialization parameters, enhancing the training stability, and preventing gradient explosion or vanishing.

[0059] The Rectified Linear Unit (ReLU) is used as the nonlinear activation function, retaining the positive features and setting the negative ones to 0. This activation function introduces nonlinear expressive power into the model. Compared with activation functions such as Sigmoid and Tanh, ReLU has no exponential operation, making it more computationally efficient. It also effectively alleviates the gradient vanishing problem in deep networks, ensuring that the feature information of the two HGTConv layers can be effectively transferred.

[0060] After ReLU activation, a Dropout layer (dropout rate of 0.2) is applied, randomly setting 20% ​​of the node features to 0. The core function of this layer is to prevent the model from overfitting, avoid the model from relying too much on the features of specific genes or brain region nodes, force the model to learn more robust heterogeneous association patterns, and improve the model's generalization ability on new datasets.

[0061] (3) Global average pooling module After being processed by two consecutive heterogeneous graph input modules, the mean of the 64-dimensional hidden features of all nodes is calculated by the global average pooling module, resulting in 64-dimensional gene modality embedding (mean of all gene node features) and 64-dimensional brain region modality embedding (mean of all brain region node features), thus realizing the aggregation from "node-level features" to "modality-level features".

[0062] (4) Modal embedding input module The modal embedding input module includes a single-connected FC1 fully connected layer, a ReLU activation layer, an FC2 fully connected layer, and a Softmax normalization layer. Its purpose is to learn dynamic fusion weights, which are then used to calculate and output global fusion features through a weighted fusion layer. Specifically: The role of the FC1 fully connected layer (the first perceptron layer) is to map the spliced ​​128-dimensional modal features (64-dimensional gene modal embeddings and 64-dimensional brain region modal embeddings) to a 32-dimensional feature space. Its core function is to perform dimensionality reduction and nonlinear transformation on the high-dimensional spliced ​​features and to mine the interaction features between modalities. The weight matrix of the fully connected layer has a dimension of 128×32. The weights are updated iteratively through gradient descent to achieve efficient compression of the feature dimension and reduce the computational complexity of subsequent operations.

[0063] After FC1, a ReLU activation layer is connected again to supplement the linear feature mapping of the fully connected layer with nonlinear expression, thereby capturing the complex nonlinear correlation between gene-image modalities.

[0064] The FC2 fully connected layer (second perceptron layer) further maps the 32-dimensional features to a 2-dimensional feature space, outputting gene modality weight scores and brain region modality weight scores. The weight matrix has a dimension of 32×2. Its core objective is to learn the importance scores of the two modalities, providing a foundation for subsequent weight normalization.

[0065] The role of the Softmax normalization layer is to normalize the 2D weight score output by FC2, thereby achieving adaptive allocation of modality weights. Compared with fixed weight fusion, it is more in line with the differences in feature contribution in different classification scenarios.

[0066] (5) Global Fusion Module The global fusion module's role is based on the dynamic weights output by Softmax:

[0067] Gene modality embedding and brain region modal embedding After weighted summation, the final fusion feature is:

[0068] The final output is a 64-dimensional globally fused feature. The core advantage of this layer is its adaptive adjustment of modality weights based on the feature distribution of the data itself, improving performance in samples with high-quality genetic features. Weighting is increased in samples with more discriminative image features. Weights are used to maximize the complementary value of multimodal features.

[0069] like Figure 6 As shown, step S5 includes a linear classifier and a Softmax activation layer. Specifically, The global fusion features are input into a linear classifier, which maps the 64-dimensional fusion features to 3-dimensional classification scores. The scores are then normalized to a classification probability distribution by a Softmax activation layer, and finally output the classification results for normal control (NC), early mild cognitive impairment (EMCI), and late mild cognitive impairment (LMCI).

[0070] The final classification is achieved by using a linear classifier and the Softmax function. The classification process is simple and efficient. Combined with feature optimization and deep fusion of multiple front-end modules, it can stably achieve accurate classification of normal cognition, early mild cognitive impairment, and late mild cognitive impairment, and has good clinical application and promotion value.

[0071] In this embodiment, the Adam optimizer and cross-entropy loss function are used to train the model, with a learning rate of 0.0005, weight decay of 1e-5, and 100 training epochs. The ten-fold cross-validation method is used to evaluate the model performance. The dataset is divided into 10 subsets, and each time 9 subsets are used as the training set and 1 subset is used as the test set. The results are repeated 10 times and the average result is taken, which effectively reduces overfitting and improves the model's generalization ability.

[0072] The results of this invention in the NC vs MCI, NC vs LMCI, NC vs EMCI, and EMCI vs LMCI classification tasks are shown in Table 1.

[0073] Table 1 Results of NC vs MCI, NC vs LMCI, NC vs EMCI, and EMCI vs LMCI classification tasks

[0074] Table 1 details the various metrics, including accuracy (ACC), sensitivity (SEN), specificity (SPE), area under the curve (AUC), and F1 score (F1). The results show that this implementation method achieves a significant improvement in accuracy, indicating that it outperforms traditional methods in classification tasks.

[0075] Example 2 like Figure 7 As shown, this embodiment discloses a classification system for mild cognitive impairment based on dynamic heterogeneous graph fusion, including: The preprocessing module is configured to preprocess single nucleotide polymorphism gene data and functional magnetic resonance imaging data to obtain standardized gene features and standardized image features. The topology-preserving dimensionality reduction module is configured to: group and reduce the dimensionality of standardized gene features, preserve the topological structure of the gene data, and output low-dimensional genetic features; The cross-modal generative completion module is configured to: based on the low-dimensional genetic features and standardized image features, complete the missing image features through a bidirectional generative adversarial network and output the completed image features; The dynamic heterogeneous graph fusion module is configured to: construct a heterogeneous graph from the low-dimensional genetic features and the completed image features, wherein the heterogeneous graph includes gene nodes and brain region nodes, and four edge types: gene-gene, brain region-brain region, gene-brain region, and brain region-gene, retaining strongly correlated edges; extract node features using a two-layer heterogeneous graph transformer convolution, input the features into a global average pooling layer to obtain gene modality embeddings and brain region modality embeddings, input the features into a perceptron to learn dynamic fusion weights, and output global fusion features; The classification module is configured to classify the global fusion features and output a classification result for mild cognitive impairment.

[0076] Example 3 The purpose of this embodiment is to provide a computer-readable storage medium.

[0077] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of a classification method for mild cognitive impairment based on dynamic heterogeneous graph fusion as described in Embodiment 1 of this disclosure.

[0078] Example 4 The purpose of this embodiment is to provide an electronic device.

[0079] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in a classification method for mild cognitive impairment based on dynamic heterogeneous graph fusion as described in Embodiment 1 of this disclosure.

[0080] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0081] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0082] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A classification method for mild cognitive impairment based on dynamic heterogeneous graph fusion, characterized in that, include: Preprocessing of single nucleotide polymorphism gene data and functional magnetic resonance imaging data yields standardized gene features and standardized image features; Standardized gene features are grouped and dimensionality reduced to preserve the topological structure of gene data and output low-dimensional genetic features. Based on the aforementioned low-dimensional genetic features and standardized image features, a bidirectional generative adversarial network is used to complete the missing image features and output the completed image features. The low-dimensional genetic features and the completed image features are used to construct a heterogeneous graph. The heterogeneous graph includes gene nodes and brain region nodes, as well as four edge types: gene-gene, brain region-brain region, gene-brain region, and brain region-gene, retaining strongly correlated edges. A two-layer heterogeneous graph transformer is used to extract node features through convolution. These features are then input into a global average pooling layer to obtain gene modality embeddings and brain region modality embeddings. The perceptron is then input to learn dynamic fusion weights, and the global fusion features are output. The global fusion features are classified to output a classification result for mild cognitive impairment.

2. The method for classifying mild cognitive impairment based on dynamic heterogeneous graph fusion as described in claim 1, characterized in that, The preprocessing steps include: Single nucleotide polymorphism (SNP) gene data preprocessing: PLINK tool was used for sample quality control, sex verification, minimum allele frequency filtering, and Hardy-Weinberg balance filtering. The data were grouped according to biological attributes, and target genes and corresponding SNP sites were selected and encoded as 0, 1, and 2. Functional magnetic resonance imaging (fMRI) image data preprocessing employed the DPABI toolkit for temporal correction, head motion correction, spatial normalization, and smoothing. Time-series features of each brain region were extracted based on brain region templates.

3. The method for classifying mild cognitive impairment based on dynamic heterogeneous graph fusion as described in claim 1, characterized in that, The grouping dimensionality reduction steps include: dividing gene features according to biological attributes, using the K-nearest neighbor local linear embedding algorithm to reduce the dimensionality of each group, and splicing the dimensionality reduction results of each group to generate low-dimensional genetic features.

4. The mild cognitive impairment classification method based on dynamic heterogeneous graph fusion as described in claim 1, characterized in that, The bidirectional generative adversarial network includes a first generator, a second generator, a first discriminator, and a second discriminator; the first generator generates completed image features based on low-dimensional genetic features, and the second generator generates verification genetic features based on real image features; the first discriminator identifies the authenticity of the completed image features, and the second discriminator identifies the authenticity of the verification genetic features. The parameters of the first and second generators are updated synchronously by calculating the adversarial loss, cross-modal topology consistency loss, and reconstruction loss.

5. The method for classifying mild cognitive impairment based on dynamic heterogeneous graph fusion as described in claim 1, characterized in that, The two-layer heterogeneous graph transformer convolution calculates the attention weights between gene nodes and brain region nodes through a type-aware attention mechanism, and then uses a feature aggregation function to weight and fuse the features of adjacent nodes to output the hidden features of the nodes.

6. The method for classifying mild cognitive impairment based on dynamic heterogeneous graph fusion as described in claim 1, characterized in that, The extracted node features are obtained by global average pooling to obtain gene modality embeddings and brain region modality embeddings. The concatenated gene modality embeddings and brain region modality embeddings are then passed through a first fully connected layer, a ReLU activation layer, and a second fully connected layer in sequence to output gene modality weight scores and brain region modality weight scores. The dynamic fusion weights are obtained by Softmax normalization. The gene modality embeddings and brain region modality embeddings are weighted and summed according to the dynamic fusion weights to output global fusion features.

7. The method for classifying mild cognitive impairment based on dynamic heterogeneous graph fusion as described in claim 1, characterized in that, A linear classifier is used to map the global fusion features to a three-dimensional classification score, which is then normalized to a classification probability distribution by a Softmax activation layer. The classification results for normal controls, early mild cognitive impairment, and late mild cognitive impairment are output.

8. A classification system for mild cognitive impairment based on dynamic heterogeneous graph fusion, characterized in that, include: The preprocessing module is configured to preprocess single nucleotide polymorphism gene data and functional magnetic resonance imaging data to obtain standardized gene features and standardized image features. The topology-preserving dimensionality reduction module is configured to: group and reduce the dimensionality of standardized gene features, preserve the topological structure of the gene data, and output low-dimensional genetic features; The cross-modal generative completion module is configured to: based on the low-dimensional genetic features and standardized image features, complete the missing image features through a bidirectional generative adversarial network and output the completed image features; The dynamic heterogeneous graph fusion module is configured to: construct a heterogeneous graph from the low-dimensional genetic features and the completed image features, wherein the heterogeneous graph includes gene nodes and brain region nodes, and four edge types: gene-gene, brain region-brain region, gene-brain region, and brain region-gene, retaining strongly correlated edges; extract node features using a two-layer heterogeneous graph transformer convolution, input the features into a global average pooling layer to obtain gene modality embeddings and brain region modality embeddings, input the features into a perceptron to learn dynamic fusion weights, and output global fusion features; The classification module is configured to classify the global fusion features and output a classification result for mild cognitive impairment.

9. A computer-readable storage medium having a program stored thereon, characterized in that, When executed by a processor, the program implements the steps in the classification method for mild cognitive impairment based on dynamic heterogeneous graph fusion as described in any one of claims 1-7.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the classification method for mild cognitive impairment based on dynamic heterogeneous graph fusion as described in any one of claims 1-7.

Images