Alzheimer disease auxiliary diagnosis method based on complementary enhanced GCN

Abstract

The invention discloses an Alzheimer's disease auxiliary diagnosis method based on complementary enhanced GCN, and relates to the technical field of medical auxiliary diagnosis. The Alzheimer's disease auxiliary diagnosis method is based on a traditional graph convolutional neural network model, inspiration of Alzheimer's disease pathogenesis is drawn, and on the premise that the advantages of the graph convolutional neural network model in the graph classification aspect are reserved, the Alzheimer's disease diagnosis efficiency is improved. Firstly, a network node and topological attribute weight self-learning unit is introduced, and the expression intensity contradiction of a traditional graph convolutional neural network on network node attribute information and topological structure information of an input signal in feature vector extraction is solved; meanwhile, a complementary enhanced GCN model fusing structural brain network and functional brain network features is constructed based on network nodes and a topological attribute weight self-learning unit, and the problem that a traditional graph convolutional neural network model based on single-type brain network data is insufficient in diagnosis accuracy, sensitivity, specificity and interpretability is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical auxiliary diagnosis, and particularly to an auxiliary diagnosis method for Alzheimer's disease based on complementary enhanced GCN. Background Art

[0002] As one of the most major neurodegenerative diseases threatening the physical and mental health of the elderly, Alzheimer's disease (AD) has attracted the common attention of the medical field and the academic field. Due to its irreversible pathological characteristics, accurate early diagnosis enables patients to obtain effective preventive and control treatments as early as possible, which is of great significance for delaying the negative development of the disease. However, the individual-differentiated clinical symptoms and limited neuroimaging accuracy have brought great challenges to the diagnosis and screening of AD. Therefore, it is particularly important to develop an efficient and reliable auxiliary diagnosis method for Alzheimer's disease, which can comprehensively improve the coverage of AD screening and diagnosis, thereby increasing the disease detection rate, improving the prognosis of patients and reducing medical costs.

[0003] Currently, many AD auxiliary diagnosis methods have been derived based on traditional graph convolutional neural networks and multimodal fusion technologies. These methods have also achieved certain improvements in diagnostic accuracy to some extent. However, many existing AD auxiliary diagnosis methods rely too much on technical-level exploration, and the exploration and integrated application of the AD pathological mechanism are quite insufficient. This not only limits the further improvement of AD auxiliary diagnosis accuracy, but also reduces the interpretability of related auxiliary diagnosis methods in the medical field, thereby affecting the future popularization and application of AD auxiliary diagnosis methods. Summary of the Invention

[0004] In order to solve the above technical problems, the present invention provides an auxiliary diagnosis method for Alzheimer's disease based on complementary enhanced GCN, including the following steps:

[0005] S1. Determine the mathematical model for predicting the diagnostic status of the subject;

[0006] S2. Define the mathematical model of the graph convolutional neural network;

[0007] S3. Determine the mathematical model of the network node and the topological attribute weight self-learning unit;

[0008] S4. Construct a complementary enhanced graph convolutional neural network model;

[0009] S5. Construct an auxiliary diagnosis system for Alzheimer's disease based on the complementary enhanced graph convolutional neural network model;

[0010] S6. Preprocess the real neuroimaging data;

[0011] S7. Train the complementary enhanced graph convolutional neural network model;

[0012] S8. Verify the performance of the Alzheimer's disease assisted diagnosis method based on the complementary enhanced graph convolutional neural network.

[0013] The further limited technical solution of the present invention is:

[0014] Further, in step S1, the structural brain network of each subject and the functional brain network are respectively denoted as and where V represents the set of network nodes in the brain network, X represents the attribute matrix of the nodes, and A s is the weighted adjacency matrix describing the anatomical connections between nodes in the structural brain network, and A f is the weighted adjacency matrix describing the functional connections between nodes in the functional brain network.

[0015] For an Alzheimer's disease assisted diagnosis method based on complementary enhanced GCN as described above, in step S1, m = |V| represents the total number of nodes in the entire brain network, and each node v i ∈V corresponds to a d-dimensional attribute vector x i , and this attribute vector is used to integrate the attribute information of the amyloid load level at each node. Then the node attribute matrix of each subject is X = [x 1 ;...; x m ∈ R m×d ; A s ∈ R m×m is determined by the strength of the white matter fiber bundles between nodes, and A f ∈ R m×m is determined by the correlation of the blood oxygenation level-dependent signals between nodes; Y = [y 1 ,..., y n represents the diagnosis status of the Alzheimer's disease conditions of all subjects, n represents the number of subjects, and the status label y i of each subject belongs to one of three clinical diagnosis states, where i = 1, 2, 3...n, and the three clinical diagnosis states are normal, mild cognitive impairment, and Alzheimer's disease.

[0016] For an Alzheimer's disease assisted diagnosis method based on complementary enhanced GCN as described above, in step S2, the mathematical model of the graph convolutional neural network is set as a graph convolutional neural network model based on the information propagation mechanism, and its function expression is defined as:

[0017]

[0018] where H( k ) represents the input of the k-th layer of the graph convolutional neural network model, and H(0 ) is the initial node attribute matrix X of the input data, i.e., H (0) = X; A ∈ R m×m represents the weighted adjacency matrix of the input data, represents the adjacency matrix with self-loops added to the network structure of the input data, where Im represents the m-order identity matrix; the degree matrix is a diagonal matrix, where the elements on the main diagonal represent the element at the i-th row and j-th column in the weighted adjacency matrix, while the elements outside the main diagonal represents a parameter matrix on the k-th layer of the graph convolutional neural network, which is shared by each node of the input data; φ(·) represents the activation function, i.e., ReLU(·) = max(0, ·).

[0019] As an Alzheimer's disease assisted diagnosis method based on complementary enhanced GCN as described above, in step S3, a network node and topological attribute weight self-learning unit is determined, and its definition is shown as follows:

[0020]

[0021] where, α (k) , β (k) respectively represent two scalar parameters on the k-th layer of the graph convolutional neural network, which are respectively used to weigh the contributions of the node attributes and topological attributes of the brain network in the representation quantity of the brain network extracted by the graph convolutional neural network.

[0022] As an Alzheimer's disease assisted diagnosis method based on complementary enhanced GCN as described above, in step S4, the complementary enhanced graph convolutional neural network model simultaneously fuses the lesion features of the structural brain network and the functional brain network. First, a two-way parallel feature extraction module is constructed using the network node and topological attribute weight self-learning unit, as shown in the following formula:

[0023]

[0024] where, and respectively represent the input quantities of the k-th layer of the improved graph convolutional neural network in the structural brain network and functional brain network pathways, and respectively represent the degree matrices of the adjacency matrices with self-loops added to the structural brain network and functional brain network; and respectively represent the scalar parameters in the k-th layer of the improved graph convolutional neural network in the structural brain network and functional brain network pathways, and respectively represent the parameter matrices in the k-th layer of the improved graph convolutional neural network in the structural brain network and the functional brain network pathway;

[0025] Then, design a feature fusion module with automatic weighted summation to superimpose the two-way feature information extracted by the feature extraction module. The definition of this process is shown in the following formula:

[0026]

[0027] where, H s and H f respectively represent the feature maps output by the structural brain network and the functional brain network pathways, and H sf represents the combined feature map output by this feature fusion module, and λ s and λ f represent two weight scalar parameters obtained through learning. represents the operation of pairwise addition of the elements at the corresponding positions in the previous and subsequent items.

[0028] As described above, in step S5 of an Alzheimer's disease auxiliary diagnosis method based on complementary enhanced GCN, the output processed by the feature fusion module in the complementary enhanced graph convolutional neural network model is sequentially connected to a multi-layer perceptron module and a Softmax unit to form an Alzheimer's disease auxiliary diagnosis system for determining the disease state of the subject.

[0029] As described above, in step S6 of an Alzheimer's disease auxiliary diagnosis method based on complementary enhanced GCN, 450 data samples are selected from the ADNI database, including data samples with three clinical diagnosis states, namely 150 normal control data samples, 150 mild cognitive impairment data samples, and 150 Alzheimer's disease data samples; each data sample has multiple T1-weighted MRI, DWI, fMRI, and Amyloid-PET scan data.

[0030] As described above, in step S6 of an Alzheimer's disease auxiliary diagnosis method based on complementary enhanced GCN, each data sample has multiple T1-weighted MRI, DWI, fMRI, and Amyloid-PET scan data, which specifically includes the following sub-steps:

[0031] S6.1. First, use the image registration method to align all other scan images with the corresponding T1-weighted MRI;

[0032] S6.2. Use the FreeSurfer software platform to perform tissue segmentation on the T1-weighted MRI images;

[0033] S6.3. Construct the cerebral cortex surface and calculate the level of Amyloid protein in each voxel;

[0034] S6.4. Use the Destrieux brain atlas to divide the cerebral cortex surface into m = 148 brain regions;

[0035] S6.5. Generate a 148×148 structural weighted adjacency matrix using surface seed-based probabilistic fiber tractography imaging technology. Each element of this matrix, i.e., the connectivity strength, is measured by the number of fibers between brain regions;

[0036] S6.6. Generate a 148×148 functional weighted adjacency matrix using a hypothesis-driven method. Each element of this matrix, i.e., the connectivity strength, is measured by the blood oxygenation level-dependent correlation between brain regions;

[0037] S6.7. Calculate the average level of Amyloid protein in each brain region;

[0038] S6.8. Select the gray matter of the cerebellum as the reference region, normalize the Amyloid protein levels in the 148 brain regions, and generate a neuropathological load of size 148×1.

[0039] In an Alzheimer's disease auxiliary diagnosis method based on complementary enhanced GCN as described above, in step S7, the preprocessed data is input into an Alzheimer's disease auxiliary diagnosis system based on a complementary enhanced graph convolutional neural network model. The model training program is written in the Python language; finally, the trained Alzheimer's disease auxiliary diagnosis system is used for the auxiliary diagnosis of the subject's disease condition.

[0040] The beneficial effects of the present invention are:

[0041] (1) In the present invention, compared with the Alzheimer's disease auxiliary diagnosis method based on the traditional graph convolutional neural network, based on the advantages of the graph convolutional neural network in processing graph structure data, inspired by the pathological mechanism of Alzheimer's disease, through the designed complementary enhanced graph convolutional neural network model, the complementary advantages of multi-modal data in improving the accuracy of Alzheimer's disease auxiliary diagnosis are skillfully exerted. It not only solves the problem of insufficient feature information of single-modal data, but also effectively solves the conflict problem caused by the simple superposition at the technical level during multi-modal data fusion;

[0042] (2) In the present invention, compared with the Alzheimer's disease assisted diagnosis method based on the traditional graph convolutional neural network, not only has a significant improvement been achieved in the screening accuracy between normal people and Alzheimer's disease patients, but also significant improvements have been achieved in the screening accuracy between normal people and patients with mild cognitive impairment, and between patients with mild cognitive impairment and Alzheimer's disease patients. This provides technical support and guarantee for the early screening and prevention of related potential patient groups. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is a schematic diagram of the overall process of the present invention;

[0044] Figure 2 is a structural diagram of the AD assisted diagnosis system based on the complementary enhanced GCN model in the embodiment of the present invention;

[0045] Figure 3 is a columnar comparison chart of the diagnostic accuracy, sensitivity, and specificity of the present invention and the traditional graph convolutional neural network among different groups. Among them, Figure (a) is a columnar comparison chart of the diagnostic accuracy, sensitivity, and specificity of the present invention and the traditional graph convolutional neural network among the NC vs. AD groups, Figure (b) is a columnar comparison chart of the diagnostic accuracy, sensitivity, and specificity of the present invention and the traditional graph convolutional neural network among the NC vs. MCI groups, and Figure (c) is a columnar comparison chart of the diagnostic accuracy, sensitivity, and specificity of the present invention and the traditional graph convolutional neural network among the MCI vs. AD groups. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0046] An Alzheimer's disease assisted diagnosis method based on complementary enhanced GCN provided in this embodiment, as Figure 1 shown, includes the following steps:

[0047] S1. Determine the mathematical model for predicting the diagnostic status of the subject; in the field of computational neuroscience, the structural brain network and the functional brain network of each subject are respectively denoted as and {V, A f , X}, where V represents the set of network nodes in the brain network, X represents the attribute matrix of the nodes, and A s is the weighted adjacency matrix describing the anatomical connections between nodes in the structural brain network, and A f is the weighted adjacency matrix describing the functional connections between nodes in the functional brain network.

[0048] Specifically, m = |V| represents the total number of nodes in the entire brain network, and each node v i ∈V corresponds to a d-dimensional attribute vector x i, this attribute vector can be used to naturally integrate the attribute information of the amyloid burden level at each node. Then, the node attribute matrix for each subject is X = [x 1 ; …; x m ∈ R m×d ; A s ∈ R m×m is determined by the strength of the white matter fiber bundles between nodes. A f ∈ R m×m is determined by the correlation of the blood oxygenation level-dependent (BOLD) signals between nodes; Y = [y 1 , …, y n represents the diagnostic status of Alzheimer's disease for all subjects. n represents the number of subjects. The status label y i of each subject belongs to one of three clinical diagnostic states, where i = 1, 2, 3... n. The three clinical diagnostic states are normal (NC), mild cognitive impairment (MCI), and Alzheimer's disease (AD).

[0049] S2. Define the mathematical model of the graph convolutional neural network; set the mathematical model of the graph convolutional neural network as a graph convolutional neural network model based on the information propagation mechanism, and its functional expression is defined as:

[0050]

[0051] Among them, H (k) represents the input of the k-th layer of the graph convolutional neural network model, and H (0) is the initial node attribute matrix X of the input data, that is, H (0) = X; A ∈ R m×m represents the weighted adjacency matrix of the input data, represents the adjacency matrix with self-loops added to the network structure of the input data, where I m represents the m-order identity matrix; the degree matrix is a diagonal matrix, where the element on the main diagonal represents the element at the i-th row and j-th column of the weighted adjacency matrix, and the elements outside the main diagonal represent a trainable parameter matrix on the k-th layer of the graph convolutional neural network, which is shared by each node of the input data; φ(·) represents the activation function, that is, ReLU(·) = max(0, ·).

[0052] S3. Determine the mathematical model of the network node and topological attribute weight self-learning unit; for the basic graph convolutional neural network model defined in step S2, as the network depth increases, the feature information of the topological structure of the input data will be more comprehensively encoded and embedded. However, the node attribute feature information of the input data will also have the problem of being over-smoothed. In view of the findings in AD pathology research: 1. AD-related amyloid proteins usually attack different brain network nodes in a certain order; 2. The development of AD pathology is accompanied by the destruction of the brain network topological structure.

[0053] In step S3, the network node and topological attribute weight self-learning unit is determined, and its definition is shown as follows:

[0054]

[0055] where α (k) , β (k) respectively represent two trainable scalar parameters of the k-th layer of the graph convolutional neural network. These two parameters can adaptively balance the contributions of the node attributes and topological attributes of the brain network in the representation of the brain network extracted by the graph convolutional neural network.

[0056] S4. Construct a complementary enhanced graph convolutional neural network model; in view of the significant damage and changes presented by AD to the topological structures of both the structural brain network and the functional brain network from different levels and angles in the current research on the AD pathogenesis mechanism, such as the reduction of neurons and synapses, the damage of white matter fibers, the reduction of neural activities, the compensation of brain functional activities, etc.

[0057] Inspired by the idea of multi-source information fusion and complementary enhancement, the complementary enhanced graph convolutional neural network model simultaneously fuses the lesion characteristics of the structural brain network and the functional brain network. As Figure 2 shown, the model first uses the network node and topological attribute weight self-learning unit to construct two parallel feature extraction modules, as shown in the following formula:

[0058]

[0059] where, and respectively represent the input quantities of the k-th layer of the improved graph convolutional neural network in the structural brain network and functional brain network pathways, and respectively represent the degree matrices of the adjacency matrices with self-loops added in the structural brain network and functional brain network; and respectively represent the trainable scalar parameters in the k-th layer of the improved graph convolutional neural network in the structural brain network and functional brain network pathways, and respectively represent the trainable parameter matrices in the k-th layer of the improved graph convolutional neural network in the structural brain network and the functional brain network pathway.

[0060] Then, a feature fusion module with automatic weighted summation is designed to flexibly superimpose the two-way feature information extracted by the feature extraction module. The definition of this process is shown in the following formula:

[0061]

[0062] where, H s and H f respectively represent the feature maps output by the structural brain network and the functional brain network pathways, H sf represents the combined feature map output by this feature fusion module, λ s and λ f represent two weight scalar parameters that can be learned, represents the operation of pairwise addition of elements at corresponding positions in the front and back terms.

[0063] S5. Construct an Alzheimer's disease auxiliary diagnosis system based on the complementary enhanced graph convolutional neural network model; connect the output processed by the feature fusion module in the complementary enhanced graph convolutional neural network model to a multi-layer perceptron module and a Softmax unit in sequence to form an Alzheimer's disease auxiliary diagnosis system for determining the disease state of the subject.

[0064] S6. Preprocess real neuroimaging data; select 450 data samples from the ADNI database, which include 150 normal control data samples (NC), 150 mild cognitive impairment data samples (MCI), and 150 Alzheimer's disease data samples (AD).

[0065] Each data sample has multiple T1-weighted MRI, DWI, fMRI, and Amyloid-PET scan data, specifically including the following sub-steps:

[0066] S6.1. First, use the image registration method to align all other scan images with the corresponding T1-weighted MRI;

[0067] S6.2. Use the FreeSurfer software platform to perform tissue segmentation on the T1-weighted MRI images;

[0068] S6.3. Construct the cerebral cortex surface and calculate the level of Amyloid protein in each voxel;

[0069] S6.4. Use the Destrieux brain atlas to divide the cerebral cortex surface into m = 148 brain regions (V);

[0070] S6.5. Generate a 148×148 structural weighted adjacency matrix (A s ) using surface-seed-based probabilistic fiber tractography imaging technology. Each element of this matrix (i.e., the connectivity strength) is measured by the number of fibers between brain regions;

[0071] S6.6. Generate a 148×148 functional weighted adjacency matrix (A f ) using a hypothesis-driven method. Each element of this matrix (i.e., the connectivity strength) is measured by the BOLD correlation between brain regions;

[0072] S6.7. Calculate the average level of Amyloid protein in each brain region;

[0073] S6.8. Select the gray matter of the cerebellum as the reference region, normalize the Amyloid protein levels of 148 brain regions, and generate a neuropathological burden (X) of size 148×1.

[0074] S7. Train a complementary enhanced graph convolutional neural network model; input the preprocessed data into the Alzheimer's disease auxiliary diagnosis system based on the complementary enhanced graph convolutional neural network model. The model training program is written in the Python language; finally, use the trained Alzheimer's disease auxiliary diagnosis system to assist in diagnosing the condition of the subjects.

[0075] S8. Verify the performance of the Alzheimer's disease auxiliary diagnosis method based on the complementary enhanced graph convolutional neural network, and conduct a comparative analysis with the traditional graph convolutional neural network in terms of the performance of disease condition diagnosis. As Figure 3 shown, the bar charts of the diagnostic accuracy (ACC), sensitivity (SEN), and specificity (SPE) of the traditional graph convolutional neural network and the proposed method in the NC vs. AD, NC vs. MCI, and MCI vs. AD groups. The type I error bar icons in the figure represent the standard deviation of the corresponding evaluation indicators.

[0076] According to Figure 3 the bar charts shown, it can be clearly seen that in the above three groups of auxiliary diagnosis tasks, the auxiliary diagnosis method proposed in this embodiment has achieved a large improvement in all three evaluation indicators; regarding the diagnostic accuracy, the smallest improvement is in the NC vs. MCI task, which also exceeds 7%; regarding the diagnostic sensitivity, the smallest improvement is in the NC vs. AD task, which also exceeds 5%; regarding the diagnostic specificity, the smallest improvement is in the NC vs. MCI task, which also exceeds 6%; Figure 3The error bars on the shown bar chart significantly indicate that the proposed auxiliary diagnosis method in this embodiment has a smaller standard deviation in the three indicators of the above three groups of tasks, thus indicating that the proposed auxiliary diagnosis method in this embodiment has stronger stability and robustness.

[0077] With the increasingly serious problem of population aging, the incidence and prevalence of Alzheimer's disease are both showing an increasing trend, which not only brings great pressure to public medical resources, but also adds a double burden of life and economy to the patients themselves and their families; at the same time, the symptoms of Alzheimer's disease have extensive heterogeneity among different patients, which undoubtedly increases the difficulty of diagnosing this disease; therefore, the emergence of the algorithm in this embodiment can effectively reduce the difficulty and cost of screening for this disease, effectively alleviate the problem of shortage of medical resources, and thus promote the harmonious development of society.

[0078] At the current medical level, as an irreversible neurodegenerative disease, early detection and timely intervention of Alzheimer's disease are still the most effective prevention and treatment means. The proposed Alzheimer's disease auxiliary diagnosis method based on complementary enhanced GCN in this embodiment is based on the traditional graph convolutional neural network model, inspired by the pathogenesis of Alzheimer's disease, and on the premise of retaining the advantages of the graph convolutional neural network model in graph classification, first introduces a network node and topological attribute weight self-learning unit to solve the contradiction between the expression intensity of the network node attribute information and topological structure information of the input signal in the traditional graph convolutional neural network when extracting feature vectors; at the same time, based on the network node and topological attribute weight self-learning unit, a complementary enhanced GCN model that integrates the characteristics of structural brain networks and functional brain networks is constructed to solve the problems of insufficient diagnostic accuracy, sensitivity, specificity, and interpretability of the traditional graph convolutional neural network model based on single-type brain network data.

[0079] In addition to the above embodiments, the present invention may also have other implementation manners. Any technical solutions formed by equivalent replacement or equivalent transformation fall within the protection scope required by the present invention.

Claims

1. A complementary enhanced GCN-based auxiliary diagnosis method for Alzheimer's disease, characterized by: The following steps are involved: S1. Determine a mathematical model for predicting the subject's diagnostic status; S2. Define the mathematical model of graph convolutional neural network; S3, determining a mathematical model of a self-learning unit for network nodes and topological attribute weights; S4, construct a complementary enhanced graph convolutional neural network model; S5. Construct an auxiliary diagnosis system for Alzheimer's disease based on a complementary enhanced graph convolutional neural network model; S6. Preprocessing real neuroimaging data; S7, training complementary enhanced graph convolutional neural network model; S8. Verify the performance of the Alzheimer's disease auxiliary diagnosis method based on complementary enhanced graph convolutional neural network.

2. The method for auxiliary diagnosis of Alzheimer's disease based on complementary enhanced GCN according to claim 1, characterized in that: In step S1, the structural brain network of each subject and functional brain networks Denoted as and Among them, V represents the set of network nodes in the brain network, X represents the attribute matrix of the node, and A s A is a weighted adjacency matrix describing the anatomical connections between nodes in a structural brain network. f A weighted adjacency matrix that describes the functional connections between nodes in a functional brain network.

3. The method for auxiliary diagnosis of Alzheimer's disease based on complementary enhanced GCN according to claim 2, characterized in that: In step S1, m=|V| represents the total number of nodes in the entire brain network, and each node v i ∈V corresponds to a d-dimensional attribute vector x i , the attribute vector is used to integrate the attribute information of the amyloid protein load level at each node, then the node attribute matrix of each subject is X = [x1; ...; x m ]∈R m×d ; A s ∈R m×m It is measured by the strength of the white matter fiber bundles between nodes, A f ∈R m×m Determined by the correlation of blood oxygen level dependence signals between nodes; Y = [y1, ..., y n ] represents the diagnosis status of Alzheimer's disease of all subjects, n represents the number of subjects, and the status label y of each subject i They all belong to one of three clinical diagnostic states, where i=1, 2, 3...n, and the three clinical diagnostic states are normal, mild cognitive impairment and Alzheimer's disease.

4. The method for auxiliary diagnosis of Alzheimer's disease based on complementary enhanced GCN according to claim 1, characterized in that: In step S2, the mathematical model of the graph convolutional neural network is set to a graph convolutional neural network model based on the information propagation mechanism, and its function expression is defined as: Among them, H (k) represents the input of the kth layer of the graph convolutional neural network model, and H (0) is the initial node attribute matrix X of the input data, that is, H (0) =X; A∈R m×m represents the weighted adjacency matrix of the input data, It represents the adjacency matrix of the input data with self-loop added to the network structure, where I m Represents the identity matrix of order m; degree matrix As a diagonal matrix, the elements on the main diagonal are represents the element at the i-th row and j-th column in the weighted adjacency matrix, while the elements outside the main diagonal i≠j; Represents a parameter matrix on the kth layer of the graph convolutional neural network, which is shared by each node of the input data; φ(·) represents the activation function, that is, ReLU(·) = max(0, ·).

5. The method for auxiliary diagnosis of Alzheimer's disease based on complementary enhanced GCN according to claim 4, characterized in that: In step S3, the network node and topology attribute weight self-learning unit is determined, and its definition is shown in the following formula: Among them, α (k) , β (k) They respectively represent two scalar parameters of the kth layer of the graph convolutional neural network, which are used to weigh the contributions of the node attributes and topological attributes of the brain network in the representation of the brain network extracted by the graph convolutional neural network.

6. The method for auxiliary diagnosis of Alzheimer's disease based on complementary enhanced GCN according to claim 5, characterized in that: In step S4, the complementary enhanced graph convolutional neural network model simultaneously integrates the structural brain network and functional brain network pathological features. First, a two-way parallel feature extraction module is constructed using the network node and topological attribute weight self-learning unit, as shown in the following formula: in, and Respectively represent the input amount of the kth layer of the improved graph convolutional neural network in the structural brain network and functional brain network pathways, and The degree matrices of the structural brain network and the functional brain network with the adjacency matrix with self-loops added are respectively; and denote the scalar parameters in the kth layer of the improved graph convolutional neural network in the structural brain network and functional brain network pathways, respectively, and W s (k) and W f (k) Represent the parameter matrices in the kth layer of the improved graph convolutional neural network in the structural brain network and functional brain network pathways respectively; Then, an automatic weighted sum feature fusion module is designed to superimpose the two feature information extracted by the feature extraction module. The definition of this process is shown in the following formula: Among them, H s and H f They represent the feature maps of the output of the structural brain network and the functional brain network pathway, respectively. sf represents the combined feature map output by the feature fusion module, λ s and λ f represents two learned weight scalar parameters, Represents the operation of adding the elements in corresponding positions of the two terms one by one.

7. The method for auxiliary diagnosis of Alzheimer's disease based on complementary enhanced GCN according to claim 6, characterized in that: In step S5, the output processed by the feature fusion module in the complementary enhanced graph convolutional neural network model is connected to the multi-layer perceptron module and the Softmax unit in sequence to form an Alzheimer's disease auxiliary diagnosis system for determining the condition of the subject.

8. The method for auxiliary diagnosis of Alzheimer's disease based on complementary enhanced GCN according to claim 1, characterized in that: In step S6, 450 data samples are selected from the ADNI database, including data samples of three clinical diagnosis states, namely, 150 normal control data samples, 150 mild cognitive impairment data samples and 150 Alzheimer's disease data samples; each data sample has multiple T1-weighted MRI, DWI, fMRI and Amyloid-PET scan data.

9. The method for auxiliary diagnosis of Alzheimer's disease based on complementary enhanced GCN according to claim 8, characterized in that: In step S6, each data sample has multiple T1-weighted MRI, DWI, fMRI and Amyloid-PET scan data, which specifically includes the following sub-steps: S6.

1. First, align all other scans with the corresponding T1-weighted MRI using an image registration method. S6.

2. Use the FreeSurfer software platform to perform tissue segmentation on T1-weighted MRI images. S6.

3. Construct the cortical surface and calculate the level of Amyloid protein in each voxel. S6.

4. Use the Destrielux brain atlas to divide the cerebral cortex surface into m = 148 brain regions; S6.5, using surface seed-based probabilistic fiber tractography to generate a 148×148 structural weighted adjacency matrix, each element of which, i.e., the connectivity strength, is measured by the number of fibers between brain regions; S6.

6. A hypothesis-driven approach was used to generate a 148×148 functional weighted adjacency matrix, in which each element, i.e., connectivity strength, was measured by the correlation of blood oxygen level dependence between brain regions. S6.

7. Calculate the average level of Amyloid protein in each brain region; S6.

8. The gray matter of the cerebellum was selected as the reference area, the Amyloid protein levels of 148 brain regions were normalized, and a neuropathological burden of size 148×1 was generated.

10. The method for auxiliary diagnosis of Alzheimer's disease based on complementary enhanced GCN according to claim 1, characterized in that: In step S7, the preprocessed data is input into the Alzheimer's disease auxiliary diagnosis system based on the complementary enhanced graph convolutional neural network model, and the model training program is written in Python language; finally, the trained Alzheimer's disease auxiliary diagnosis system is used to perform auxiliary diagnosis of the subject's condition.

Images