Cerebral cortex wrinkle classification system based on multi-scale map fusion hierarchical map convolution method

Abstract

The invention discloses a cerebral cortex wrinkle classification system based on a multi-scale atlas fusion hierarchical graph convolution method. Firstly, a group of well-defined multi-scale atlas is utilized to calculate a multi-scale functional connection network; then, on the basis of the biological hierarchical relationship among the regions in the multi-scale atlas, node pool operation is implemented on multiple spatial scales, and'atlas guided pooling 'is established; the characteristics of the cerebral cortex wrinkles are comprehensively extracted from a multi-scale function connection network by combining a hierarchical graph convolution technology and a'graph guided pooling 'method, so that effective classification of the cerebral cortex wrinkles is realized.

Description

Technical Field

[0001] The present invention belongs to the cross-technical field of medical imaging and neural networks, and specifically relates to a cerebral cortex fold classification system based on a multi-scale atlas fusion layered graph convolution method. Background Art

[0002] The brain is the most important organ of the human body. Understanding the structure and function of the brain is one of the most challenging frontier scientific issues in the 21st century. The folded structure of the cerebral cortex is an important manifestation of functional complexity. The sulci and gyri are two important anatomical structures of the cerebral cortex. By analyzing the function and structure of cortical folds, we can better understand the functions of sulci and gyri.

[0003] With the development of neuroimaging technology, functional magnetic resonance imaging (fMRI) and diffusion tensor imaging (DTI) can observe the structure and functional connections of brain sulci and gyri in more detail. Through these technologies, more accurate brain connection maps can be drawn, further revealing the complex relationship between the structure and function of cerebral cortical folds.

[0004] In previous methods, for functional mapping analysis of cortical folds, fast Fourier transform is often used to calculate the power spectrum of gyrus and sulcus signals, so as to directly classify cortical folds. With the rapid development of deep learning technology in medical images, convolutional neural networks are also used for classification. A classifier is designed based on a one-dimensional convolutional neural network to distinguish the fMRI signals of gyri and sulci in each brain region.

[0005] However, most of the current classification methods of cortical folds directly use functional signals to classify sulci and gyri. In these studies, although spatiotemporal graph convolutional networks provide a powerful tool to reveal the complex connection patterns of the cerebral cortex, there are still problems. For example, there is a lack of a comprehensive understanding of the complex topological relationships and functional differences across the whole brain. Specifically, existing studies mainly focus on the analysis of specific regions and tasks, and pay less attention to the functional connectivity of the whole brain. Moreover, most of the current studies on cortical folds use a single-scale method, which ignores the differences in the functional mapping of cortical folds at different spatial resolutions. For the functional mapping of cortical folds, it is not completely accurate to classify cortical folds simply by relying on functional signals. Therefore, research in this area has produced a new perspective, allowing us to have a deeper understanding of the complexity and diversity of the functional mapping of cortical folds. Summary of the invention

[0006] The purpose of the present invention is to provide a cerebral cortical fold classification system based on a multi-scale atlas fusion layered graph convolution method, so as to solve the problem in the prior art of ignoring the differences in the functional mapping of cortical folds at different spatial resolutions, and the problem that the classification of cerebral cortical folds relying solely on functional signals is not completely accurate.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is:

[0008] A cerebral cortical fold classification system based on multi-scale atlas fusion layered graph convolution method, the specific steps are as follows:

[0009] Step 1: Acquire resting state fMRI data and preprocess them to obtain a four-dimensional fMRI voxel image, and convert the four-dimensional fMRI voxel image into a one-dimensional fMRI vector;

[0010] Step 2: Acquire diffusion MRI data and preprocess them to obtain DTI data, acquire structural MRI (T1-weighted) data and calculate the curvature of each voxel;

[0011] Step 3: Division of multi-scale atlas, using the DTI data in step 2, the reconstructed cortical surface is divided, all voxel points on the brain surface are spectrally clustered, and the brain atlas is divided into 442 regions (ROIs), 364 regions (ROIs), 274 regions (ROIs), and 189 regions (ROIs);

[0012] Step 4: The curvature of each voxel point calculated from the structural MRI data in step 2 is used. According to the set threshold, high curvature values ​​are defined as sulci, and low curvature values ​​are defined as gyri. If more than half of the points in a region are located in the sulci, then the region is marked as the sulcus region, and the remaining regions are all gyri regions, which are used as labels for regions of interest (ROIs);

[0013] Step 5: Match the one-dimensional fMRI vector in step 1 to the four templates divided in step 3, using The Pearson correlation coefficient between the BOLD time series of two regions of interest (ROIs) was calculated as the functional connectivity between the ROIs; where v i and u j Respectively represent the observed values ​​of voxel point i and voxel point j in the sample set, corr(v i , v j ) is the voxel point v i With voxel point v j The correlation between cov(v i , v j ) is the voxel point v i With voxel point v jThe covariance between i and σ j Represents voxel point v i With voxel point u j The standard deviation of

[0014] Step 6: Using the multi-scale atlas from step 3, we construct an atlas from fine-scale segmentation by merging the fine-scale segmented ROIs into a larger coarse-scale segmented ROI. (with R ROIs) to the coarse-scale segmentation atlas The mapping matrix M (with P ROIs) R-→p , and R>P;

[0015] Step 7: Construct a four-layer hierarchical graph convolutional network of multi-scale atlas. Specifically, the hierarchical graph convolutional network of multi-scale atlas performs graph convolution between node features and FCN in the finest segmentation (i.e., 442-ROI atlas) until processing the FCN in the coarsest scale atlas (i.e., 189-ROI atlas), using the multi-scale atlas divided in step 3 and the functional connectivity matrix calculated in step 5;

[0016] Step 8: Use the functional connectivity matrix under the 189-ROI map in step 7, perform dimensionality reduction on it for classification, and then use the cross entropy loss function to optimize the model.

[0017] Further, the specific steps of step 6 are as follows:

[0018] Step a: Calculate the number of overlapping voxels for fine-scale maps For each ROIi in the coarse-scale atlas, calculate its The number of spatially overlapping voxels of each ROIj in , and the number of voxels that co-exist between ROIi and ROIj are counted;

[0019] Step b: Calculate the overlap ratio and further calculate the overlap ratio It reflects the coverage of fine-scale ROI in coarse-scale ROI;

[0020] Step c: Set the overlap ratio ρ ij > 0, then in the mapping matrix M R→P Mark 1, i.e. M R→P [i, j]=1.

[0021] Further, the specific steps of step 7 are as follows:

[0022] Step a: The structure of the graph can be effectively represented by the adjacency matrix A and the degree matrix D. The adjacency matrix A is the functional connection matrix calculated in step 5, and the degree matrix D uses a one-hot encoding feature, namely the identity matrix, using the spectral convolution idea. Perform graph convolution operation; is the normalized adjacency matrix, H (l) is the node feature matrix of the lth layer, W (l) is a learnable weight matrix, σ is an activation function;

[0023] Step b: Use the mapping matrix in step 6 as “graph-guided pooling” to connect graph convolutional networks of different scales, and output the feature vector h GC442 (a 442×1 vector)h AP364 (a 364×1 vector) and used with FCN in the next layer of graph convolution;

[0024] Step c: Layered graph convolution and atlas-guided pooling are continued until the FCN in the coarsest scale atlas (i.e., 189-ROI atlas) is processed.

[0025] Compared with the prior art, the present invention has the following technical features:

[0026] First, the functional connection matrix is ​​used to classify the folds of the cerebral cortex. This is different from the existing technology that relies on functional signals to classify the folds of the cerebral cortex, which cannot be completely accurate. Second, the idea of ​​multi-scale atlas layering captures and analyzes spatial features through different scales and layers, taking into account the difference in the functional mapping of cortical folds at different spatial resolutions. Third, layered graph convolution is introduced to capture the features of different levels in the graph. Compared with traditional methods, layered graph convolution processes complex graph structures through a layered approach, enhancing the expressiveness and flexibility of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is the basic structure diagram of the graph convolutional network (GCN).

[0028] Figure 2 This is an overview of the multi-scale graph fusion layered graph convolutional network. DETAILED DESCRIPTION

[0029] The present invention is further described below with reference to the embodiments and the accompanying drawings.

[0030] A method for classifying cerebral cortical folds based on a multi-scale atlas fusion layered graph convolution method. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0031] In a first aspect, the present invention provides a multi-scale atlas layering concept, including:

[0032] The idea of ​​multi-scale atlas layering is to capture and analyze spatial features through different scales and levels. Different scales represent different spatial resolutions and feature levels. Extracting features at different scales can provide a more comprehensive understanding of the data. Combined with machine learning and deep learning methods, multi-scale features are used for classification tasks.

[0033] The brain can be analyzed at multiple levels, by dividing it into different numbers of regions of interest (ROIs), and by analyzing the strength of connections at different levels of the brain through multi-scale connectomics methods, including tracing of white matter fiber tracts and construction of functional connectivity networks.

[0034] Combining different types of data, functional magnetic resonance imaging obtains functional connectivity networks, and diffusion tensor imaging calculates the curvature of each vertex, thereby obtaining the corresponding labels of sulci and gyri, which can provide a more comprehensive perspective of brain networks. Through the idea of ​​multi-scale atlas layering, the functional and structural characteristics of the brain at different levels are revealed.

[0035] Multiscale brain networks are not simply a replication of single-scale structures, but are essentially different aspects of the brain system. Brain network connections at different scales have different topological properties, and multiscale hierarchical structures are of great significance for studying brain cognition. Multiscale hierarchical brain networks support switching between the integration and separation of information to achieve efficient brain function. Moreover, cerebral cortical folds present different functional connection network topologies at specific or multiple scales.

[0036] In a second aspect, the present invention provides a method for classifying cerebral cortex folds by connecting different layers in series using a graph convolutional network, comprising:

[0037] Graph Convolutional Network (GCN) is a deep learning model for graph data. By performing convolution operations on the graph structure, it can effectively capture the relationship between nodes and the global features of the graph.

[0038] Hierarchical Graph Convolutional Network (HGCN) is a deep learning model for graph data, which captures features at different levels in the graph through a hierarchical structure. Compared with traditional graph convolutional networks (GCN), HGCN processes complex graph structures through a hierarchical approach, enhancing the expressiveness and flexibility of the model.

[0039] In the present invention, the functional connectivity matrix of the brain is used as a graph, which consists of nodes and edges. The nodes represent regions of interest (ROIs), and the edges are the functional connections between ROIs. The representation of each node is updated by aggregating the features of the nodes and using the information of the neighboring nodes. HGCN divides the graph into multiple levels or subgraphs according to the features of the nodes. The output of each layer is obtained by weighted summing the node features of the current layer and the features of the neighboring nodes.

[0040] In each layer, the features of the nodes not only depend on the direct neighbors, but also aggregate information from different levels through the hierarchical structure. By aggregating features from different levels, HGCN can capture multi-scale information, thereby improving the performance of the model.

[0041] A cerebral cortex fold classification system based on a multi-scale atlas fusion layered graph convolution method, the system comprising a processor and a memory for storing a plurality of functional modules that can be run on the processor, the functional modules comprising an acquisition module, a preprocessing module, a layered graph convolution module and a classification module;

[0042] The acquisition module is used to obtain resting state fMRI data;

[0043] The preprocessing module is used to preprocess the resting state fMRI data to obtain a four-dimensional fMRI voxel image, convert the four-dimensional fMRI voxel image into a one-dimensional fMRI vector, preprocess the diffusion MRI data to obtain DTI data, and preprocess the structural MRI (T1-weighted) data to calculate the curvature of each voxel;

[0044] The layered graph convolution module is used to generate a four-layer layered graph convolution network of a multi-scale graph. Specifically, the layered graph convolution network of the multi-scale graph performs graph convolution between the node features in the finest segmentation (i.e., the 442-ROI graph) and the FCN until processing the FCN in the coarsest scale graph (i.e., the 189-ROI graph).

[0045] Among them, the multi-scale atlas layered graph convolution layer has the following steps:

[0046] Step a: The structure of the graph can be effectively represented by the adjacency matrix A and the degree matrix D. The adjacency matrix A is the functional connection matrix, and the degree matrix D uses a one-hot encoding feature, namely the identity matrix, using the spectral convolution idea. Perform graph convolution operation; is the normalized adjacency matrix, H (l) is the node feature matrix of the lth layer, W (l) is a learnable weight matrix, σ is an activation function;

[0047] Step b: Use the mapping matrix as “graph-guided pooling” to connect graph convolutional networks of different scales, and output the feature vector h GC442 (a 442×1 vector)h AP364 (a 364×1 vector) and is used with the FCN in the next layer of graph convolution.

[0048] Step c: Layered graph convolution and atlas-guided pooling are continued until the FCN in the coarsest scale atlas (i.e., 189-ROI atlas) is processed.

[0049] The classification module is used to perform dimensionality reduction on the output results of the layered graph convolution module for classification, and then use the cross entropy loss function to optimize the model.

[0050] Furthermore, the preprocessing module includes the following submodules:

[0051] The first submodule is used to preprocess the resting state fMRI data, wherein the preprocessing includes skull removal, head motion correction, slice time correction, spatial smoothing, linear trend removal and bandpass filtering to obtain a four-dimensional fMRI voxel image;

[0052] The second submodule is used to obtain a one-dimensional fMRI vector by cropping and resampling the T1 and fMRI data, linearly registering them, and spatially transforming them into the DTI space;

[0053] The third submodule is used to connect the time points of all one-dimensional fMRI vectors to obtain a two-dimensional fMRI matrix.

[0054] In summary, the present invention proposes a hierarchical GCN method based on multi-scale atlases, which uses the prior brain network hierarchy under the GCN framework to capture the differences in the functional connection network of cerebral cortical folds more effectively and sensitively. First, a multi-scale atlas is implemented to construct a multi-scale functional connection network to avoid potential biases in single-scale analysis and comprehensively characterize the brain network. A map-guided pooling is further invented to perform node aggregation based on the prior hierarchical relationships defined in the multi-scale atlas. Therefore, a hierarchical structure is constructed by stacking GCNs and map-guided pooling layers to extract features related to cerebral cortical folds for final classification.

[0055] First, the technical terms appearing in the present invention are explained:

[0056] Resting-state fMRI: Resting-state fMRI is a method of collecting data while the subject is in a resting state and not performing any specific tasks.

[0057] FCN: A method to reveal how the brain works together at rest or during a specific task by analyzing the functional connectivity between different brain regions.

[0058] In this embodiment, a method for classifying cerebral cortex wrinkles based on a multi-scale atlas fusion layered graph convolution method is disclosed. Figure 2 As shown, it includes layered graph convolution, graph-guided pooling, fully connected dimensionality reduction, and classification;

[0059] The steps include:

[0060] Step 1: Acquire resting state fMRI data and preprocess them to obtain a four-dimensional fMRI voxel image, and convert the four-dimensional fMRI voxel image into a one-dimensional fMRI vector;

[0061] Step 2: Acquire diffusion MRI data and preprocess them to obtain DTI data, acquire structural MRI (T1-weighted) data and calculate the curvature of each voxel;

[0062] Step 3: Division of multi-scale atlas. No existing template is used. The DTI data in step 2 is used to divide the reconstructed cortical surface. All voxel points on the brain surface are spectrally clustered to divide the brain atlas. The brain is divided into 442 regions (ROIs), 364 regions (ROIs), 274 regions (ROIs), and 189 regions (ROIs). The areas of the divided regions are basically equal and are used as regions of interest (ROIs) for the next step of the experiment.

[0063] Step 4: The curvature of each voxel point calculated from the structural MRI data in step 2 is used. According to the set threshold, high curvature values ​​are defined as sulci, and low curvature values ​​are defined as gyri. If more than half of the points in a region are located in the sulci, then the region is marked as the sulcus region, and the remaining regions are all gyri regions, which are used as labels for regions of interest (ROIs);

[0064] Step 5: Match the one-dimensional fMRI vector in step 1 to the four templates divided in step 3. For each subject, the whole brain is divided into 442ROIs, 364ROIs, 274ROIs, and 189ROIs, and represents a network with 442 / 364 / 274 / 189 nodes. First, calculate the average fMRI signal of each brain region and normalize it by standard Z-score normalization, the formula is: where f i is the average fMRI signal of brain region i, f μ and f σ is the mean and standard deviation of all fMRI signals, and then, using The Pearson correlation coefficient between the BOLD time series of two regions of interest (ROIs) was calculated as the functional connectivity between the ROIs; where v i and v j Respectively represent the observed values ​​of voxel point i and voxel point j in the sample set, corr(v i , v j ) is the voxel point v i With voxel point v j The correlation between cov(v i , v j ) is the voxel point v iWith voxel point v j The covariance between i and σ j Represents voxel point v i With voxel point v j The standard deviation of

[0065] Step 6: Using the multi-scale atlas from step 3, we construct an atlas from fine-scale segmentation by merging the fine-scale segmented ROIs into a larger coarse-scale segmented ROI. (with R ROIs) to the coarse-scale segmentation atlas The mapping matrix M (with P ROIs) R→P , and R>P, so as to use the spatial overlap of regions of interest (ROIs) of different scales to represent the hierarchical relationship of different scales. The specific steps are as follows:

[0066] Step a: Calculate the number of overlapping voxels for fine-scale maps For each ROIi in the coarse-scale atlas, calculate its The number of spatially overlapping voxels of each ROIj in . Count the number of voxels that exist in common between ROIi and ROIj;

[0067] Step b: Calculate the overlap ratio and further calculate the overlap ratio It reflects the coverage of fine-scale ROI in coarse-scale ROI;

[0068] Step c: Set the overlap ratio ρ ij > 0, then in the mapping matrix M R→P Mark 1, i.e. M R→P [i, j] = 1, and then the fine-scale features are summed by weighted summation. Mapping to coarse-scale features

[0069] Step 7: Construct a four-layer hierarchical graph convolutional network of multi-scale atlas. Specifically, the hierarchical graph convolutional network of multi-scale atlas performs graph convolution between node features in the finest segmentation (i.e., 442-ROI atlas) and FCN until processing the FCN in the coarsest scale atlas (i.e., 189-ROI atlas). The multi-scale atlas divided in step 3 and the functional connection matrix calculated in step 5 are used. The specific steps are as follows:

[0070] Step a: Figure 2As shown in the figure, spectral graph convolution is used to construct a graph convolution network. The core idea is to extract node features through spectral graph convolution, so as to achieve effective representation and learning of the graph. The structure of the graph can be effectively represented by the adjacency matrix A and the degree matrix D. The adjacency matrix A is the functional connection matrix calculated in step 5, and the degree matrix D uses a one-hot encoding feature, that is, the unit matrix. The spectral convolution can be expressed as: Where gk(L) is the polynomial function of the Laplace operator, H(l) is the node feature matrix of the lth layer, and W (l) is a learnable weight matrix, and σ is an activation function. To simplify the calculation, GCNs usually adopt a first-order approximation and transform it into: Perform graph convolution operation; is the normalized adjacency matrix, H (l) is the node feature matrix of the lth layer, W (l) is a learnable weight matrix, σ is an activation function;

[0071] Step b: Use the mapping matrix in step 6 as “graph-guided pooling” to connect graph convolutional networks of different scales, and output the feature vector h GC442 (a 442×1 vector)h AP364 (a 364×1 vector) and is used with the FCN in the next layer of graph convolution.

[0072] Step c: Layered graph convolution and atlas-guided pooling are continued until the FCN in the coarsest scale atlas (i.e., 189-ROI atlas) is processed.

[0073] Step 8: Use the functional connectivity matrix under the 189-ROI map in step 7, perform dimensionality reduction on it for classification, and then use the cross entropy loss function to optimize the model.

[0074] The functional connection matrix is ​​input for training separately, and the cross entropy loss function is used to calculate the training loss. Each iteration records the current global optimal network structure, and replaces the original optimal network during the iteration process. Finally, after reaching the maximum number of iterations, the optimal model with the smallest neural network loss will be selected.

[0075] In this embodiment, a cerebral cortex fold classification system based on a multi-scale atlas fusion layered graph convolution method is also disclosed, the system includes a processor and a memory for storing multiple functional modules that can be run on the processor, the functional modules include an acquisition module, a preprocessing module, a layered graph convolution module and a classification module;

[0076] The acquisition module is used to obtain resting state fMRI data;

[0077] The preprocessing module is used to preprocess the resting state fMRI data to obtain a four-dimensional fMRI voxel image, convert the four-dimensional fMRI voxel image into a one-dimensional fMRI vector, preprocess the diffusion MRI data to obtain DTI data, and preprocess the structural MRI (T1-weighted) data to calculate the curvature of each voxel;

[0078] The layered graph convolution module is used to generate a four-layer layered graph convolution network of a multi-scale graph. Specifically, the layered graph convolution network of the multi-scale graph performs graph convolution between the node features in the finest segmentation (i.e., the 442-ROI graph) and the FCN until processing the FCN in the coarsest scale graph (i.e., the 189-ROI graph).

[0079] Among them, the multi-scale atlas layered graph convolution layer has the following steps:

[0080] Step a: The structure of the graph can be effectively represented by the adjacency matrix A and the degree matrix D. The adjacency matrix A is the functional connection matrix, and the degree matrix D uses a one-hot encoding feature, namely the identity matrix, using the spectral convolution idea. Perform graph convolution operation; is the normalized adjacency matrix, H (l) is the node feature matrix of the lth layer, W (l) is a learnable weight matrix, σ is an activation function;

[0081] Step b: Use the mapping matrix as “graph-guided pooling” to connect graph convolutional networks of different scales, and output the feature vector h GC442 (a 442×1 vector)h AP364 (a 364×1 vector) and is used with the FCN in the next layer of graph convolution.

[0082] Step c: Layered graph convolution and atlas-guided pooling are continued until the FCN in the coarsest scale atlas (i.e., 189-ROI atlas) is processed.

[0083] The classification module is used to perform dimensionality reduction on the output results of the layered graph convolution module for classification, and then use the cross entropy loss function to optimize the model.

[0084] Furthermore, the preprocessing module includes the following submodules:

[0085] The first submodule is used to preprocess the resting state fMRI data, wherein the preprocessing includes skull removal, head motion correction, slice time correction, spatial smoothing, linear trend removal and bandpass filtering to obtain a four-dimensional fMRI voxel image;

[0086] The second submodule is used to register T1 and fMRI to DTI space through FLIRT in FM-RIB Software Library (FSL) to obtain a one-dimensional fMRI vector;

[0087] The third submodule is used to connect the time points of all one-dimensional fMRI vectors to obtain a two-dimensional fMRI matrix.

[0088] This example performs 100 iterations of training, with the learning rate set to 0.001 and the batch size set to 30. Adam is used as the optimizer, and the weight decay is set to 0.1 to avoid overfitting.

[0089] In order to evaluate the performance of the model, the data set was randomly divided into a training set and a test set with a ratio of 80% and 20%. The accuracy (ACC), sensitivity (SEN), specificity (SPE) and the area under the receiver operating characteristic curve (area under the ROC curve, AUC) were selected as performance indicators.

[0090] Table 2 Experimental results

[0091]

Claims

1. A cerebral cortical fold classification system based on multi-scale atlas fusion layered graph convolution method, characterized in that: The specific steps are as follows: Step 1: Acquire resting state fMRI data and preprocess them to obtain a four-dimensional fMRI voxel image, and convert the four-dimensional fMRI voxel image into a one-dimensional fMRI vector; Step 2: Acquire diffusion MRI data and preprocess them to obtain DTI data, acquire structural MRI (T1-weighted) data and calculate the curvature of each voxel; Step 3: Division of multi-scale atlas, using the DTI data in step 2, the reconstructed cortical surface is divided, all voxel points on the brain surface are spectrally clustered, and the brain atlas is divided into 442 regions (ROIs), 364 regions (ROIs), 274 regions (ROIs), and 189 regions (ROIs); Step 4: The curvature of each voxel point calculated from the structural MRI data in step 2 is used. According to the set threshold, high curvature values ​​are defined as sulci, and low curvature values ​​are defined as gyri. If more than half of the points in a region are located in the sulci, then the region is marked as the sulcus region, and the remaining regions are all gyri regions, which are used as labels for regions of interest (ROIs); Step 5: Match the one-dimensional fMRI vector in step 1 to the four templates divided in step 3, using The Pearson correlation coefficient between the BOLD time series of two regions of interest (ROIs) was calculated as the functional connectivity between the ROIs; Among them, v i and v j Respectively represent the observed values ​​of voxel point i and voxel point j in the sample set, corr(v i ,v j ) is the voxel point v i With voxel point v j The correlation between cov(v i ,v j ) is the voxel point v i With voxel point v j The covariance between i and σ j Represents voxel point v i With voxel point v j The standard deviation of Step 6: Using the multi-scale atlas from step 3, we construct an atlas from fine-scale segmentation by merging the fine-scale segmented ROIs into a larger coarse-scale segmented ROI. (with R ROIs) to the coarse-scale segmentation atlas The mapping matrix M (with P ROIs) R→P , and R>P; Step 7: Construct a four-layer hierarchical graph convolutional network of multi-scale atlas. Specifically, the hierarchical graph convolutional network of multi-scale atlas performs graph convolution between node features and FCN in the finest segmentation (i.e., 442-ROI atlas) until processing the FCN in the coarsest scale atlas (i.e., 189-ROI atlas), using the multi-scale atlas divided in step 3 and the functional connectivity matrix calculated in step 5; Step 8: Use the functional connectivity matrix under the 189-ROI map in step 7, perform dimensionality reduction on it for classification, and then use the cross entropy loss function to optimize the model.

2. According to claim 1, a cerebral cortex fold classification system based on multi-scale atlas fusion layered graph convolution method is characterized in that: Step 6 The specific steps are as follows: Step a: Calculate the number of overlapping voxels for fine-scale maps For each ROIi in the coarse-scale atlas, calculate its The number of spatially overlapping voxels of each ROIj in , and the number of voxels that co-exist between ROIi and ROIj are counted; Step b: Calculate the overlap ratio and further calculate the overlap ratio It reflects the coverage of fine-scale ROI in coarse-scale ROI; Step c: Set the overlap ratio ρ ij >0, then in the mapping matrix M R→P Mark 1, i.e. M R→P [i,j]=1.

3. According to claim 1, a cerebral cortex fold classification system based on multi-scale atlas fusion layered graph convolution method is characterized in that: The specific steps of step 7 are as follows: Step a: The structure of the graph can be effectively represented by the adjacency matrix A and the degree matrix D. The adjacency matrix A is the functional connection matrix calculated in step 5, and the degree matrix D uses a one-hot encoding feature, namely the identity matrix, using the spectral convolution idea. Perform graph convolution operation; is the normalized adjacency matrix, H (l) is the node feature matrix of the lth layer, W (l) is a learnable weight matrix, σ is an activation function; Step b: Use the mapping matrix in step 6 as "graph-guided pooling" to connect graph convolutional networks of different scales, and output the feature vector h GC442 (a 442×1 vector)h AP364 (a 364×1 vector) and used with FCN in the next layer of graph convolution; Step c: Layered graph convolution and atlas-guided pooling are continued until the FCN in the coarsest scale atlas (i.e., 189-ROI atlas) is processed.

Images