A diagnostic method for autism spectrum disorder incorporating weighted learning networks

Abstract

This invention relates to a diagnostic method for autism spectrum disorder (ASD) incorporating a weighted learning network. The weighted learning network represents the similarity of non-imaging data in the latent space, providing a novel method for constructing edge weights in a population graph. Defining pairwise associations in the latent space rather than the input space is beneficial and robust. This invention proposes a residual connection method using graph convolutional neural networks (GCNs). By introducing residual units, it mitigates information loss caused by convolution operations, avoiding the vanishing and exploding gradient problems. Simultaneously, the EdgeDrop strategy, by randomly removing edges from the original graph, makes node connections sparser, mitigating overfitting and oversmoothing issues that occur during DeepGCN training. Experiments show that the application of data augmentation algorithms improves model accuracy, feature extraction, and classification performance. The DeepGCN classification method incorporating a weighted learning network proposed in this invention can accurately identify ASD and exhibits excellent model performance.

Description

Technical fields:

[0001] This invention relates to the interdisciplinary field of medicine and artificial intelligence, specifically to the automatic diagnosis of autism spectrum disorder using deep learning technology, and more particularly to a DeepGCN method that combines a weighted learning network. Background technology:

[0002] Autism spectrum disorder (ASD) is a neurodevelopmental disorder characterized by social impairment, repetitive and stereotyped behaviors, and restricted interests, and its prevalence is increasing year by year. ASD is a chronic disease with a poor prognosis; most children with ASD are unable to live independently in adulthood and require lifelong care. Currently, the diagnosis of ASD mainly relies on clinical characteristics. Doctors determine whether a patient has ASD based on their behavior and daily performance, a method that is highly subjective. Therefore, there is an urgent need for an objective diagnostic method to assist doctors in diagnosing patients.

[0003] With the development of medical imaging technology and artificial intelligence, machine learning-based methods have achieved excellent results in analyzing and studying the brain magnetic resonance imaging (MRI) of patients with autism spectrum disorder (ASD). In recent years, a large amount of non-imaging datasets have been collected. Factors such as patients' gene sequences, gender, and IQ play a crucial role in disease diagnosis. However, due to differences in scanning sites, equipment, and parameters among patients, a significant amount of uncontrollable heterogeneity is hidden behind the data. Currently, most classification algorithms focus only on single-modal data. For disease diagnosis classification tasks, this means focusing solely on imaging data, while the use of non-imaging data is insufficient. Just as clinicians diagnose diseases not only based on imaging but also on a comprehensive assessment of factors such as the patient's height, weight, and medical history, the same applies to using artificial intelligence to diagnose ASD. The complementary use of multimodal imaging and non-imaging data will greatly contribute to improving the performance of classification algorithms. However, traditional machine learning methods cannot effectively utilize non-imaging data.

[0004] Currently, deep learning methods offer the possibility of diagnosing mental illnesses through multimodal data fusion. However, general deep learning methods are not suitable for multimodal modeling, and the model representation is not direct, resulting in unsatisfactory model performance.

[0005] In recent years, graph neural networks have been proposed to address this problem. Graph Convolutional Networks (GCNs) extend convolution operations on Euclidean data to non-Euclidean graph data. GCN updates originate from a combination of local node-level context and global neighborhood-level context. Graph Laplacian is used to aggregate neighborhood information of nodes in the graph. GCNs leverage their structural advantages to utilize complex medical multimodal and multiscale data to extract valuable information hidden within them.

[0006] This invention proposes DeepGCN with a weighted learning network for the diagnosis of ASD. First, we propose a weighted learning network to construct the edge weights of the graph, which can fully exploit the correlations of non-imaging data (e.g., age, gender, IQ) in the latent space. Simultaneously, based on the idea of ​​ResNet in CNNs, residual connections of graph convolutional neural networks are proposed to solve the gradient problem of DeepGCN. Finally, we introduce the EdgeDrop strategy to randomly discard some edges before training, making the original graph sparser and thus reducing the overly smooth convergence speed. The DeepGCN combined with the weighted learning network proposed in this invention for diagnosing autism spectrum disorder exhibits excellent model performance. Summary of the Invention:

[0007] The purpose of this invention is to overcome the shortcomings of existing methods in fully utilizing multimodal data and improve model performance. It proposes a DeepGCN classification method to address the problems of low accuracy, low precision, low recognition standard scores, and poor model performance in autism spectrum disorder identification models.

[0008] A diagnostic method for autism spectrum disorder combining a weighted learning network is proposed. This method utilizes a deep Generative Network (GCN) for feature learning and classification. Imaging information is used to represent node features in the population graph, and phenotypic information is used to construct edges in the population graph. The weighted learning network obtains edge weights from pairwise associations of phenotypic information in the latent space. Node prediction results are clustered using a semi-supervised learning method. The steps are as follows:

[0009] Step 1: Select a suitable dataset and preprocess it through the preprocessing pipeline:

[0010] Step 1.1: Exclude some subject data from the autism spectrum disorder dataset that lack time series data, have incomplete brain coverage, severe head movement, artifacts, and other scanning artifacts, and finally retain the appropriate data;

[0011] Step 1.2: Select a spectrum and preprocess it using the preprocessing channel;

[0012] Step 1.3: First, the rs-fMRI data were corrected for slice time and motion, and the voxel intensity was normalized;

[0013] Step 1.4: Smooth the functional image using a 5mm full width at half height (FWHM) Gaussian kernel to remove interference signals;

[0014] Step 1.5: The functional data is time-filtered using a bandpass filter of 0.01-0.1Hz. Spatial registration is then performed in the template space (MNI152) using a nonlinear method.

[0015] Step 2: Construct nodes for the population graph using functional magnetic resonance imaging data, and reduce the dimensionality of the original features using recursive feature elimination;

[0016] Step 2.1: Calculate the time series;

[0017] Step 2.2: Calculate the feature matrix of the preprocessed data based on the Pearson correlation coefficient;

[0018] Step 2.3: Use Fisher z-transform to improve the consistency of variance, thereby obtaining the whole-brain functional connectivity matrix for each topic;

[0019] Step 2.4: The values ​​in the upper triangle of the matrix were removed, as they were duplicates of the values ​​in the lower triangle, as well as the main diagonal of the matrix, which represents a measure of correlation with itself;

[0020] Step 2.5: Use Recursive Feature Elimination (RFE) to reduce the dimensionality of the feature vectors. In each iteration, the Ridge classifier is trained on the training set using the current feature vectors. The coefficients of the classifier are used to rank the importance of the features, and the feature vectors with the smallest coefficients are discarded. This process is repeated until a fixed number of features are obtained.

[0021] Step 3: Construct the edges of the population graph using non-imaging phenotypic data, and obtain the edge weights using a weighted learning network;

[0022] Step 3.1: Construct the edges of the population graph using phenotypic data;

[0023] Step 3.2: Normalize and rescale the non-imaging input data u to a unit norm vector.

[0024] Step 3.3: The rescaled vector passes through three fully connected layers with output dimensions h1, h2, and h2, respectively. Except for the last layer, each fully connected layer is followed by a LeakyReLU activation layer, a batch normalization layer, and a dropout layer.

[0025] Step 3.4: Calculate the edge weight w(i,j) between the two nodes. We will input the two non-imaging inputs u i Perform the above operations on both uj and uj, sharing the same weights, and finally calculate the cosine similarity between the two latent feature vectors. The cosine similarity is calculated using the following formula:

[0026]

[0027] Step 4: Use DeepGCN based on spectral graph convolution as a classifier;

[0028] Step 4.1: Define the graph Fourier transform by analogy with the Euclidean domain using the Laplace eigenfunction. The formula for the traditional Fourier transform is shown below:

[0029] f(x)=∫f(t)e -i2πxt dt

[0030] The important part of the above formula is e. -i2πxt In this equation, i represents the complex field, t represents time, x represents the frequency, and 2πxt represents the rotation angle within time t, which plays a significant role in the Fourier transform. In fact, e -i2πxt It is the generalized eigenfunction of the Laplacian operator Δ. For an image, its Laplacian operator is actually a Laplacian matrix L, and the eigenvector u corresponding to this matrix represents e. -i2πxt Specifically, the Laplace matrix can be represented as:

[0031] L=DA

[0032] That is, the degree matrix (degree of each point) minus the adjacency matrix (weights between two points). Normalizing the Laplacian matrix yields:

[0033]

[0034] The Laplace matrix is ​​a positive semi-definite matrix with real symmetry, allowing for spectral decomposition, and the diagonal elements of Λ are always non-negative. Further eigendecomposition of L yields:

[0035] L=UΛU T

[0036] Where U=(u1,u2,...,u n This is the characteristic matrix, and the eigenvectors U of Laplace are... T Using the eigenvectors U of the Laplace matrix as the basis for the Fourier transform, performing a Fourier transform on the graph is equivalent to applying the eigenvectors U of the Laplace matrix to the graph. T It can be represented by a linear combination;

[0037] Step 4.2: Performing a Fourier transform on the graph is equivalent to applying the Laplace matrix eigenvectors U to the graph. T The linear combination of these terms is used to represent the features x of a node in a graph. Therefore, for a feature x, the traditional Fourier formula is transformed into a graph-based Fourier formula:

[0038]

[0039] The inverse Fourier transform is Ux. Therefore, according to the convolution theorem, the traditional convolution operation can be transformed into a convolution operation formula located on a graph:

[0040] (x*g)(t)=F-1[F[x(t)]⊙F[g(t)]]=U((UTx)(UTg))

[0041] Where ⊙ represents the Hadamard product. We use a learnable convolution kernel gθ to represent UTG, then the Hadamard product can be represented in the form of matrix multiplication. The spectral convolution is represented as follows: UgθUTx

[0042] Step 4.3: To improve computational efficiency and focus on the neighborhood information of nodes, and because the above spectral convolution is computationally intensive, a k-order Chebyshev polynomial is used to approximate the convolution kernel:

[0043]

[0044] Where T k (·) is a Chebyshev polynomial of order k, β k These are the corresponding coefficients. It is a rescaled eigenvalue diagonal matrix. By incorporating matrix operations into Chebyshev polynomials, we can obtain the convolution operation of one layer of GCN:

[0045]

[0046] Where σ is the activation function, and the final convolution operation is defined as:

[0047]

[0048] in This eliminates the need for eigenvalue decomposition of the Laplace matrix during calculation. In actual computation, the properties of Chebyshev polynomials can be used for recursion.

[0049]

[0050] also, Convolutional neural networks (CNNs) can be used to design and train deep models, thereby better representing features and improving model performance. The concept of Deep Graph Convolutional Neural Networks (DeepGCN) was thus proposed. DeepGCN is composed of multiple stacked convolutional layers, and the relationship between the layers is as follows:

[0051]

[0052] y (l) It is the feature representation of the l-th layer, y (0) =x, W (l) σ is the learnable parameter of the l-th layer, and σ is the activation function. It is an adjacency matrix The degree matrix;

[0053] Step 5: Propose residual connections in GCN to avoid the problems of vanishing and exploding gradients during training;

[0054] Step 5.1: Convert the solved mapping function H(x) into the network's residual mapping function H(x) = F(x) + x. The output of this layer and the output of the previous layer together determine the input of the next layer.

[0055] Step 5.2: Update the weight values ​​of the F(x) part to get infinitely close to the optimal solution;

[0056] Step 5.3: Input the completed graph into G. h Each of the four convolutional layers contains a ReLU activation layer. Except for the last convolutional layer (output layer), each layer also contains a dropout layer. The output layer does not use residual units.

[0057] Step 5.4: New Feature Representation The input to the MLP is P, which represents the number of hidden and output layer units.

[0058] Step 6: Propose the EdgeDrop strategy to alleviate the problems of overfitting and oversmoothing during training;

[0059] Step 6.1: Randomly select E from the adjacency matrix A during training. q Set a non-zero element to zero;

[0060] Step 6.2: After discarding, the new adjacency matrix is ​​obtained and used as the final input of DeepGCN;

[0061] Step 7: Based on DeepGCN, complete the classification and diagnosis of autism spectrum disorder using the methods described above.

[0062] Compared with the prior art, the beneficial effects of the present invention are:

[0063] This invention provides a diagnostic method for autism spectrum disorder that combines weighted learning networks. By using deep GCN to represent features and cluster nodes, the supplementary information from non-imaging data helps in node classification, which is beneficial to improving model performance and has broader application prospects.

[0064] By using a weight learning network to learn the edge weights of the population graph in GCN, the results of random weight initialization are effective. The weight learning network fully learns the pairwise associations between non-imaging phenotypic data, which is beneficial to improving the classification performance of the model.

[0065] In addition, this invention designs GCN residual connections for use in deep GCNs. This connection method combines the output of the current layer and the output of the previous layer as the input of the lower layer, which alleviates information loss and effectively avoids the problems of gradient vanishing and gradient explosion during training. The GCN residual connections proposed in this invention can be extended to any graph neural network.

[0066] Finally, this invention designs the EdgeDrop strategy, which performs edge discarding operation on the constructed population graph, randomly discarding some edges, increasing the sparsity of the population graph, and effectively alleviating the overfitting and oversmoothing problems that occur in DeepGCN during training.

[0067] Experiments using the HO atlas on the ABIDE-I dataset show that the method of this invention significantly outperforms current ASD classification methods, achieving state-of-the-art performance, which demonstrates the effectiveness of our method. Attached image description:

[0068] Figure 1 This is an overall flowchart of a diagnostic method for autism spectrum disorder that incorporates weighted learning networks;

[0069] Figure 2 This is a structural diagram of a diagnostic method for autism spectrum disorder that incorporates a weighted learning network;

[0070] Figure 3 This is a model framework diagram of a weighted learning network;

[0071] Figure 4 This is a flowchart of the classification process using graph convolutional neural network residual connections in DeepGCN;

[0072] Figure 5 This is a simulation diagram of the present invention;

[0073] Figure 6 It is a visualization of the original features and the features after classification. Detailed implementation method:

[0074] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0075] Figure 1 The diagram illustrates the overall process of implementing this invention. Figure 2 The diagram shows the model structure proposed in this invention. The DeepGCN method, which combines a weighted learning network, for diagnosing autism spectrum disorder includes the following steps:

[0076] Step 1: Select a suitable dataset and preprocess it using the Configurable Pipeline for the Analysis of Connectomes (C-PAC):

[0077] Step 1.1: Data from the autism spectrum disorder dataset was excluded due to lack of time series data, incomplete brain coverage, severe head movement, artifacts, and other scanning artifacts. Finally, data from 871 subjects were used, including 403 individuals with ASD and 468 typical controls.

[0078] Step 1.2: Select the Harvard-Oxford Cortical Structural Map (HO) and preprocess it using channels in C-PAC. The HO generates 111 Regions of Interest (ROIs).

[0079] Step 1.3: First, the rs-fMRI data were corrected for slice time and motion, and the voxel intensity was normalized;

[0080] Step 1.4: Smooth the functional image using a 5mm full width at half height (FWHM) Gaussian kernel to remove interference signals;

[0081] Step 1.5: The functional data is time-filtered using a bandpass filter of 0.01-0.1Hz. Spatial registration is then performed in the template space (MNI152) using a nonlinear method.

[0082] Step 2: Construct nodes for the population graph using functional magnetic resonance imaging data, and reduce the dimensionality of the original features using recursive feature elimination;

[0083] Step 2.1: Calculate the time series between each pair of ROIs;

[0084] Step 2.2: Calculate the feature matrix of the preprocessed data based on the Pearson correlation coefficient. This matrix effectively measures the linear correlation between two time series u and v of a given length T. The formula is as follows:

[0085]

[0086] in, and Represents the average time series;

[0087] Step 2.3: Use Fisher z-transform to improve the consistency of variance, thereby obtaining the whole-brain functional connectivity matrix for each topic;

[0088] Step 2.4: The values ​​of the upper triangle in the matrix were deleted because they were repeated from the values ​​of the lower triangle, as well as the main diagonal of the matrix, which represents the correlation measure with itself. Finally, only the strict lower triangle part was retained to represent the functional connectivity between any two ROIs. The final retained matrix was vectorized and expanded into a one-dimensional vector to obtain the 6105-dimensional feature vector required for classification.

[0089] Step 2.5: Use Recursive Feature Elimination (RFE) to reduce the dimensionality of the feature vectors. In each iteration, the Ridge classifier is trained on the training set using the current feature vectors. The coefficients of the classifier are used to rank the importance of the features, and the feature vectors with the smallest coefficients are discarded. This process is repeated until a fixed number of 2000 features are obtained.

[0090] Figure 3 The diagram shows the model framework of the weight learning network proposed in this invention;

[0091] Step 3: Construct the edges of the population graph using non-imaging phenotypic data, and obtain the edge weights using a weighted learning network;

[0092] Step 3.1: Construct the edges of the population graph using phenotypic data (including site, age, and gender);

[0093] Step 3.2: Normalize and rescale the non-imaging input data u to a unit norm vector.

[0094] Step 3.3: The rescaled vector passes through three fully connected layers with output dimensions h1=256, h2=128, and h3=128 respectively. Except for the last layer, each fully connected layer is followed by a LeakyReLU activation layer, a batch normalization layer, and a dropout layer.

[0095] Step 3.4: Calculate the edge weight w(i,j) between two nodes. We perform steps 3.1, 3.2, and 3.3 on the two non-imaging inputs ui and uj respectively, sharing the weights. Finally, we calculate the cosine similarity of the two latent feature vectors. The cosine similarity is calculated using the following formula:

[0096]

[0097] Where v i and v j Is it an imaging input u? i and u j The latent feature vector obtained after processing by the WL network;

[0098] Figure 4 This is a flowchart of the classification process using graph convolutional neural network residual connections in DeepGCN;

[0099] Step 4: Use DeepGCN based on spectral graph convolution as a classifier;

[0100] Step 4.1: Define the graph Fourier transform by analogy with the Euclidean domain using the Laplace eigenfunction. The formula for the traditional Fourier transform is shown below:

[0101] f(x)=∫f(t)e -i2πxt dt

[0102] The important part of the above formula is e. -i2πxt In this equation, i represents the complex field, t represents time, x represents the frequency, and 2πxt represents the rotation angle within time t, which plays a significant role in the Fourier transform. In fact, e -i2πxt It is the generalized eigenfunction of the Laplacian operator Δ. For an image, its Laplacian operator is actually a Laplacian matrix L, and the eigenvector u corresponding to this matrix represents e. -i2πxt Specifically, the Laplace matrix can be represented as:

[0103] L=DA

[0104] That is, the degree matrix (degree of each point) minus the adjacency matrix (weights between two points). Normalizing the Laplacian matrix yields:

[0105]

[0106] The Laplace matrix is ​​a positive semi-definite matrix with real symmetry, allowing for spectral decomposition, and the diagonal elements of Λ are always non-negative. Further eigendecomposition of L yields:

[0107] L=UΛU T

[0108] Where U=(u1,u2,...,u n This is the characteristic matrix, and the eigenvectors U of Laplace are... T Using the eigenvectors U of the Laplace matrix as the basis for the Fourier transform, performing a Fourier transform on the graph is equivalent to applying the eigenvectors U of the Laplace matrix to the graph. T It can be represented by a linear combination;

[0109] Step 4.2: Performing a Fourier transform on the graph is equivalent to applying the Laplace matrix eigenvectors U to the graph. T The linear combination of these terms is used to represent the features x of a node in a graph. Therefore, for a feature x, the traditional Fourier formula is transformed into a graph-based Fourier formula:

[0110]

[0111] The inverse Fourier transform is Ux. Therefore, according to the convolution theorem, the traditional convolution operation can be transformed into a convolution operation formula located on a graph:

[0112] (x*g)(t)=F-1[F[x(t)]⊙F[g(t)]]=U((UTx)⊙(UTg))

[0113] Where ⊙ represents the Hadamard product. We use a learnable convolution kernel gθ to represent UTG, then the Hadamard product can be represented in the form of matrix multiplication, and the spectral convolution is represented as follows:

[0114] Ug θ U T x

[0115] Step 4.3: To improve computational efficiency and focus on the neighborhood information of nodes, and because the above spectral convolution is computationally intensive, a k=3rd order Chebyshev polynomial is used to approximate the convolution kernel:

[0116]

[0117] Where T k (·) is a Chebyshev polynomial of order k=3, β k These are the corresponding coefficients. It is a rescaled eigenvalue diagonal matrix. By incorporating matrix operations into Chebyshev polynomials, we can obtain the convolution operation of one layer of GCN:

[0118]

[0119] Where σ is the ReLU activation function, and the final convolution operation is defined as:

[0120]

[0121] in This eliminates the need for eigenvalue decomposition of the Laplace matrix during calculation. In actual computation, the properties of Chebyshev polynomials can be used for recursion.

[0122]

[0123] also, Convolutional neural networks (CNNs) can be used to design and train deep models, thereby better representing features and improving model performance. The concept of Deep Graph Convolutional Neural Networks (DeepGCN) was thus proposed. DeepGCN is composed of multiple stacked convolutional layers, and the relationship between the layers is as follows:

[0124]

[0125] y (l)It is the feature representation of the l-th layer, y (0) =x, W (l) Here, σ is the learnable parameter of the l-th layer, and σ is the ReLU activation function. It is an adjacency matrix The degree matrix;

[0126] Step 5: Propose residual connections in GCN to avoid the problems of vanishing and exploding gradients during training;

[0127] Step 5.1: Convert the solved mapping function H(x) into the network's residual mapping function H(x) = F(x) + x. The output of this layer and the output of the previous layer together determine the input of the next layer.

[0128] Step 5.2: Update the weight values ​​of the F(x) part to get infinitely close to the optimal solution;

[0129] Step 5.3: Input the completed graph into G. h =8 convolutional layers, each followed by a ReLU activation layer. Except for the last convolutional layer (output layer), each layer is followed by a dropout layer. The output layer does not use residual units.

[0130] Step 5.4: New Feature Representation The input to the MLP is P=128, which represents the number of hidden and output layer units.

[0131] Step 6: Propose the EdgeDrop strategy to alleviate the problems of overfitting and oversmoothing during training;

[0132] Step 6.1: Randomly select E from the adjacency matrix A during training. q = Set 30% of non-zero elements to zero;

[0133] Step 6.2: After discarding, the new adjacency matrix is ​​obtained and used as the final input of DeepGCN;

[0134] Step 7: Based on DeepGCN, complete the classification and diagnosis of autism spectrum disorder using the methods described above.

[0135] The effects of the present invention will be further described below with reference to simulation experiments.

[0136] 1. Simulation conditions:

[0137] The hardware test platform for the simulation experiment of this invention is: CPU is E5-2680 v4 with a main frequency of 2.4GHz, GPU is NVIDIA Tesla V100 with 32G of video memory; the software platform is Ubuntu 18.04.

[0138] 2. Simulation content and result analysis:

[0139] The simulation experiment of this invention uses the method of this invention to train and test the autism spectrum disorder classification model on ABIDE-I, the world's largest public dataset for autism spectrum disorder. This invention uses multimodal imaging data and non-imaging data, performs feature representation and classification through DeepGCN, and tests the model performance.

[0140] To evaluate the performance of the autism spectrum disorder classification model obtained in this invention, a 10-fold cross-validation method was used, and accuracy (ACC), precision (Precision), recall, F1 score, and area under the ROC curve (AUC) were used as evaluation metrics. Figure 5 (a) shows the box plots for various performance metrics under 10-fold cross-validation. Figure 5 (b) is the ROC curve under the test set.

[0141] from Figure 5 As can be seen from (a), the best-case performance metrics of this invention are close to 1. Precision reflects the depth of the model's predictions, i.e., the proportion of samples predicted as ASD that are actually ASD. Our method exhibits an average precision of 88.0%, indicating high model confidence. Furthermore, the average recall of 87.6% reflects high model breadth, i.e., the proportion of samples that were actually ASD that the model successfully predicted. The F1 score also achieved an excellent result of 87.0%. Figure 5 As can be clearly seen in (b), the area under the curve of the model is large, and the average AUC is 0.87, which indicates that the model has very high robustness. In addition, the False Positive Rate increases at the beginning of the ROC curve, which means that the model has a higher diagnostic rate for ASD and a lower misdiagnosis rate, making it possible to provide accurate and reliable results for clinical applications.

[0142] See the attached diagram in the instruction manual. Figure 6 The visualization results show that the two types of original features are randomly distributed in the binary mapping, and the nodes of the two types are mixed and disordered, without clear boundaries. This reflects the great difficulty of classifying directly using the original features; the algorithm clearly distinguishes ASD patients from typical controls and reflects the differences between categories. This indicates that the fused multimodal features exhibit better intra-class clustering performance and inter-class discriminability, proving the effectiveness of the invention.

[0143] It should be understood that any parts not described in detail in this specification belong to the prior art.

[0144] The above description, in conjunction with the accompanying drawings, is merely a specific embodiment and process of the present invention. However, the scope of protection of the present invention is not limited thereto. Any person skilled in the art should understand that this is only an illustrative example, and various changes and substitutions can be made to this embodiment without departing from the essence of the present invention. The scope of the present invention is defined only by the appended claims.

[0145] The embodiments described with reference to the accompanying drawings are exemplary and are only used to explain the invention. They should not be construed as limiting the invention. The specific scope of the embodiments of the invention is not limited thereto. On the contrary, all embodiments of the invention include all variations and modifications that fall within the spirit and scope of the appended claims.

Claims

1. A diagnostic method for autism spectrum disorder incorporating a weighted learning network, characterized in that, Deep Generative Networks (GCNs) are used for feature learning and classification. Imaging information is used to represent node features in the population graph, and phenotypic information is used to construct edges in the population graph. The weight learning network obtains edge weights from pairwise associations of phenotypic information in the latent space. The node prediction results are clustered using a semi-supervised learning method. The steps are as follows: Step 1: Select a suitable dataset and preprocess the dataset through the preprocessing pipeline; Step 1.1: Exclude some subject data from the autism spectrum disorder dataset that lack time series data, have incomplete brain coverage, severe head movement, artifacts, and other scanning artifacts, and finally retain the appropriate resting-state functional magnetic resonance imaging (rs-fMRI) data. Step 1.2: For the retained rs-fMRI data, select an atlas and preprocess it using a preprocessing channel, the preprocessing including steps 1.3 to 1.5; Step 1.3: First, the rs-fMRI data were corrected for slice time and motion, and the voxel intensity was normalized; Step 1.4: The data after Step 1.3 is smoothed using a 5 mm full width at half height (FWHM) Gaussian kernel to remove interference signals; Step 1.5: The data after step 1.4 is time-filtered using a bandpass filter of 0.01-0.1Hz, and the filtered data is registered to the MNI152 standard brain template space using a nonlinear method; Step 2: Construct the nodes of the population graph using the preprocessed data obtained in Step 1, and use recursive feature elimination to reduce the dimensionality of the original features; Step 2.1: Using the preprocessed data obtained in Step 1.5 as the object, determine multiple regions of interest based on the spectrum selected in Step 1.2, calculate the average signal value of all voxels in each region of interest at each time point, and obtain the average time series of each region of interest; Step 2.2: Calculate the feature matrix for all average time series data based on the Pearson correlation coefficient; Step 2.3: Apply Fisher's z-transform to the feature matrix to improve the consistency of variance, thereby obtaining the whole-brain functional connectivity matrix for each topic; Step 2.4: The values ​​of the upper triangle in the matrix are deleted, which are repeated in the lower triangle, as well as the main diagonal of the matrix, which represents the correlation measure with itself. The lower triangle elements are expanded into a one-dimensional vector in sequence to obtain the original functional connectivity feature vector of each subject. Step 2.5: Use Recursive Feature Elimination (RFE) to reduce the dimensionality of the original feature vectors. In each iteration, the Ridge classifier is trained on the training set using the current feature vectors. The coefficients of the classifier are used to rank the importance of the features, and the feature vectors with the smallest coefficients are discarded. This process is repeated until a fixed number of node feature vectors are obtained. Step 3: Construct the edges of the population graph using non-imaging phenotypic data, and obtain the edge weights using a weighted learning network; Step 3.1: Construct the edges of the population graph using non-imaging phenotypic data; Step 3.2: Non-imaging phenotypic data Normalize and rescale to a unit norm vector ; Step 3.3: The rescaled vector passes through three fully connected layers, with output dimensions of respectively... , , Except for the last layer, each fully connected layer is followed by a LeakyReLU activation layer, a batch normalization layer, and a dropout layer. Step 3.4: Place the first Non-imaging phenotypic data of individual subjects and the Non-imaging phenotypic data of individual subjects Input the three fully connected layers described in step 3.3, while sharing weights, to obtain the first... Latent feature vectors corresponding to subjects and the Latent feature vectors corresponding to each subject Then, the cosine similarity of the latent feature vectors is calculated, and the edge weight between the two nodes is obtained based on the cosine similarity. : Step 4: Use DeepGCN based on spectral graph convolution as a classifier. The feature vectors of each subject obtained in Step 2.5 are used as the node features of the population graph, and the edge weights between nodes obtained in Step 3.4 are used as the edge weights in the adjacency matrix of the population graph. The population graph input to DeepGCN is constructed. DeepGCN performs graph convolution feature learning on the population graph and outputs the deep feature representation of each subject node for subsequent classification and diagnosis. Step 4.1: Define the graph Fourier transform by analogy with the Euclidean domain using the Laplace eigenfunction. The formula for the traditional Fourier transform is shown below: The important part of the above formula is... ,in Represents the field of complex numbers. It indicates time. It represents frequency. It represents time. The rotation angle inside plays a relatively important role in the Fourier transform, among which... It is the Laplace operator The generalized eigenfunction of an image is, for example, its Laplacian operator is simply a Laplacian matrix. The eigenvectors corresponding to this matrix This means Specifically, the Laplace matrix can be represented as: Subtracting the adjacency matrix from the degree matrix and normalizing the Laplace matrix yields: The Laplace matrix is ​​a positive semi-definite matrix and possesses real symmetry, enabling spectral decomposition. Diagonal elements must be non-negative. Further eigenvalue decomposition yields: in It's the eigenma matrix, the eigenvectors of Laplace. Using the eigenvectors of the Laplace matrix as the basis for the Fourier transform, performing a Fourier transform on the graph is equivalent to applying the eigenvectors of the Laplace matrix to the graph. It can be represented by a linear combination; Step 4.2: Performing a Fourier transform on the graph is equivalent to applying the Laplace matrix eigenvectors to the graph. It is represented by a linear combination of features, therefore, for the features of nodes in the graph The traditional Fourier formula is transformed into a graph-based Fourier formula: Inverse Fourier transform Therefore, according to the convolution theorem, the traditional convolution operation can be transformed into a convolution operation formula located on a graph: in Representing the Hadamard product, we use a learnable convolutional kernel. to indicate Therefore, the Hadamard product can be represented in the form of matrix multiplication, and the spectral convolution is represented in the following form: Step 4.3: To improve computational efficiency and focus on the neighborhood information of nodes, and because the above spectral convolution computation is computationally intensive, we use... Chebyshev polynomials are used to approximate the convolution kernel: in yes Chebyshev polynomials These are the corresponding coefficients. It is a rescaled eigenvalue diagonal matrix. By incorporating matrix operations into Chebyshev polynomials, we can obtain the convolution operation of one layer of GCN: in As the activation function, the final convolution operation is defined as: in This eliminates the need for eigenvalue decomposition of the Laplace matrix during calculation. In actual computation, the properties of Chebyshev polynomials can be used for recursion. also, Convolutional neural networks (CNNs) can be used to design and train deep models, thereby better representing features and improving model performance. The concept of Deep Graph Convolutional Neural Networks (DeepGCN) was thus proposed. DeepGCN is composed of multiple stacked convolutional layers, and the relationship between the layers is as follows: It is the first Layer feature representation, , It is the first Learnable parameters of the layer For activation function, It is an adjacency matrix The degree matrix; Step 5: Set up residual connections in the graph convolutional neural network in Step 4 to avoid the problems of gradient vanishing and gradient exploding during training; Step 5.1: Solve the mapping function Residual mapping function converted to network ,in, This represents the input features of the current convolutional layer. This indicates the current convolutional layer's response to the input features. The residual mapping obtained after graph convolution operation This represents the output feature of the current convolutional layer after adding residual connections. The output of this layer and the output of the previous layer together determine the input of the next layer. Step 5.2: Update Some of the weight values ​​are infinitely close to the optimal solution; Step 5.3: Input the completed graph into... Each convolutional layer contains a ReLU activation layer, and except for the last convolutional layer, each layer also contains a dropout layer. The last convolutional layer is the output layer of DeepGCN, used to output the final node feature representation of DeepGCN. The output layer does not use residual units; Step 5.4: New Feature Representation Input into MLP This refers to the number of hidden layer and output layer units; Step 6: Propose the EdgeDrop strategy to alleviate the problems of overfitting and oversmoothing during training; Step 6.1: Randomly select edge weights during training. middle Set a non-zero element to zero; Step 6.2: After discarding, the new adjacency matrix is ​​obtained and used as the final input of DeepGCN; Step 7: Based on DeepGCN, complete the classification and diagnosis of autism spectrum disorder using the methods described above.

2. The method for diagnosing autism spectrum disorder by incorporating a weighted learning network according to claim 1, characterized in that, The weight learning network constructed in step 3 consists of three fully connected layers with output dimensions of 256, 128, and 128, respectively. Except for the last layer, each fully connected layer is followed by a LeakyReLU activation layer, a batch normalization layer, and a dropout layer. The cosine similarity is calculated using the following formula: in and Is it an imaging input? and The latent feature vectors obtained after processing by the WL network.

Images