Cognitive-related brain disease classification apparatus, device, medium, and product

Abstract

The application provides a cognitive-related brain disease classification device, equipment, medium and product. First, an rs-fMRI image is obtained, and then the rs-fMRI image is converted into a BFC graph structure represented by nodes and edges. The scale label is included in the node, and a plurality of heterogeneous types are included in the edge. Then, a trained graph learning neural network is used to process the BFC graph to obtain a classification result. In this way, the diversified features of the hierarchical heterogeneous BFC graph derived from the multi-space scale brain template are obtained based on the early fusion strategy, the brain function connection mode in different cognitive-related brain disease stages can be better understood, and more accurate cognitive-related brain disease classification can be achieved.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a device, equipment, medium and product for classifying cognitive-related brain diseases. Background Technology

[0002] With the development of neuroimaging and artificial intelligence technologies, many machine learning algorithms, especially deep learning algorithms, based on magnetic resonance imaging (MRI) data have been proposed for analyzing whole-brain connectomes, simulating abnormalities in the human brain connectome, or clinically distinguishing patients from normal controls. In particular, thanks to the unique advantages of graph neural networks (GNNs) in brain network data analysis, many existing works use GNNs to analyze brain functional connectivity (BFC) maps obtained from resting-state functional MRI (rs-fMRI) data, thereby classifying and diagnosing phenotypic states of cognitive-related brain diseases such as Alzheimer's disease. To mitigate the spatial representation bias caused by constructing BFC maps based on a single segmented brain template, recent work focuses on developing novel large-scale BFC map learning methods based on multiple brain templates to integrate complementary network topological information from multiple BFC maps.

[0003] However, existing methods have three shortcomings: (1) Existing methods mostly use completely different brain segmentation templates to divide brain regions, and these templates have almost no spatial correlation (such as the commonly used AAL template, Harvard-Oxford template, CC200 template, etc.). Existing methods and techniques cannot model the hierarchical and multi-scale brain functional connectivity structure of the human brain; (2) Existing studies usually assume that the connections between brain regions defined at different atlas scales are homogeneous. The few multi-scale BFC map learning methods proposed recently only focus on the heterogeneity of connections within the same scale and between spatially overlapping cross-scale brain regions. Existing methods ignore the importance of the functional patterns implied by the connections between non-spatially overlapping cross-scale brain regions for the representation of different stages of cognitive-related brain diseases; (3) The representation fusion methods used in existing methods mainly rely on late fusion strategies, such as averaging or splicing the features of multiple BFC maps before classification. This intuitive late fusion strategy cannot fully capture the complex spatial dependencies between different brain maps in the early stages of cognitive-related brain diseases.

[0004] The aforementioned problems with current methods and technologies result in insufficient accuracy for existing classification and diagnostic tasks for cognitive-related brain diseases, failing to meet clinical needs. Summary of the Invention

[0005] This application proposes a device, equipment, medium, and product for classifying cognitive-related brain diseases, which can solve one of the problems existing in the prior art.

[0006] To achieve the above objectives, this application adopts the following technical solution: In a first aspect, a device for classifying cognitive-related brain diseases is provided, comprising: The acquisition unit is used to acquire resting-state functional magnetic resonance imaging (rs-fMRI) images of the brain of the subjects to be processed. The initial processing unit is used to process the rs-fMRI images of the subject to be processed to obtain the brain functional connectivity (BFC) diagram structure of the subject. ,in, , , For the subjects, express The set of nodes in the middle, express Adjacency matrix, scale label Nodes representing a specific spatial scale The corresponding Region of Interest (ROI) is defined for each edge. , express The corresponding heterogeneous types include: same-scale edges, cross-scale overlapping edges, and cross-scale non-overlapping edges; and, The classification unit is used to classify the BFC diagram structure of the subject to be processed. The input is fed into a trained graph learning neural network for processing, and the corresponding classification result is obtained.

[0007] Based on the above technical solution, rs-fMRI images are first obtained, and then the rs-fMRI images are converted into BFC graph structures represented by nodes and edges. The nodes contain scale labels, and the edges contain several heterogeneous types. Then, a trained graph learning neural network is used to process the BFC graph to obtain classification results. In this way, by obtaining the diverse features of hierarchical heterogeneous BFC graphs derived from multi-spatial-scale brain templates based on the early fusion strategy, we can better understand the brain functional connectivity patterns at different stages of cognitive-related brain diseases and achieve more accurate classification of cognitive-related brain diseases.

[0008] In one possible design of the first aspect, the initial processing unit specifically includes: The preprocessing unit is used to preprocess the rs-fMRI images of the subject to be processed to obtain the brain oxygenation level-dependent BOLD signal sequence of the subject to be processed; and, The construction unit is used to construct a multi-scale hierarchical BFC map of the subject's brain using brain partition templates at different spatial scales; calculate the average value of the subject's BOLD signal within voxels of ROIs in the multi-scale hierarchical BFC map; concatenate the average values ​​of the subject's BOLD signals to form a cross-scale time series matrix; and construct the initial BFC map structure of the subject based on the cross-scale time series matrix. ,in, Indicates the subject BFC diagram, express The set of nodes in the middle, express The adjacency matrix, The number of subjects is indicated; and the initial BFC graph structure of the subjects to be processed is expanded to the BFC graph structure of the subjects to be processed.

[0009] In one possible design of the first aspect, the weights of the edges of the initial graph structure of the BFC of the subject to be processed are represented by the Pearson correlation coefficient, Spearman correlation coefficient, partial correlation function or cross-correlation function of the average time series signal of the edges.

[0010] In one possible design of the first aspect, the preprocessing includes: time point correction, spatial smoothing, head motion correction, noise removal, spatial normalization, and / or bandpass filtering.

[0011] In one possible design approach of the first aspect, the same-scale edge is defined as: for a graph Any two nodes in and If and only if and Coming from the same spatial scale When, its corresponding edge Defined as edges of the same scale and having ; The cross-scale overlapping edge is defined as follows: given two different map scales... Two sets of ROI and Assuming Compare With finer spatial granularity and more ROIs defined, then for ROI of any one of them voxel set and Any ROI A collection of voxels, and Spatial overlap between Defined as: in, This represents the number of voxels within a specific ROI, if Then the edge Defined as and The cross-scale overlapping hierarchical edges and have , To adjust the parameters; The cross-scale non-overlapping edge is defined as: given two different map scales... Two sets of ROI and ,for ROI of any one of them voxel set and Any ROI A set of voxels if and only if At that time, then the side Defined as and Non-overlapping edges between scales and have .

[0012] In one possible design of the first aspect, the classification unit specifically includes: The heterogeneous subgraph feature aggregation unit is used to process, for any node in the BFC graph structure of the subject to be processed, the subject's same-scale subgraph corresponding to the same-scale edge using a graph convolutional neural network to obtain the subject's same-scale feature embedding; to aggregate the subject's cross-scale overlapping subgraph corresponding to the cross-scale overlapping edge using a hierarchical pooling mechanism to obtain the subject's cross-scale overlapping feature embedding; to aggregate the subject's cross-scale non-overlapping subgraph corresponding to the cross-scale non-overlapping edge using heterogeneous graph convolution operations to obtain the subject's cross-scale non-overlapping feature embedding; to concatenate the subject's cross-scale overlapping feature embedding and the subject's cross-scale non-overlapping feature embedding to obtain the subject's cross-scale feature embedding; and to aggregate the subject's same-scale feature embedding and the subject's cross-scale feature embedding to obtain the subject's node embedding features. A full-image feature fusion unit is configured to perform full-scale fusion of the embedded features of all nodes in the BFC graph structure of the subject to be processed, thereby obtaining the full-image fused features of the subject to be processed; and... The classification subunit is used to process the full-image fusion features of the subject to be processed to obtain the classification result.

[0013] In one possible design approach of the first aspect, the classification unit employs the cross-entropy loss function Lce: in, Indicates the number of samples, and These represent the true label and the predicted label, respectively.

[0014] In a second aspect, an electronic device is provided, comprising: a processor, and a memory coupled to the processor, the memory being used to store a computer program; the processor being used to execute the computer program stored in the memory, such that the electronic device performs the following process: Obtain resting-state functional magnetic resonance imaging (rs-fMRI) images of the brains of the subjects to be treated; The rs-fMRI images of the subjects to be treated were processed to obtain the brain functional connectivity (BFC) diagram structure of the subjects. ,in, , , For the subjects, express The set of nodes in the middle, express Adjacency matrix, scale label Nodes representing a specific spatial scale The corresponding Region of Interest (ROI) is defined for each edge. , express The corresponding heterogeneous types include: same-scale edges, cross-scale overlapping edges, and cross-scale non-overlapping edges; and, The BFC diagram structure of the subject to be processed The input is fed into a trained graph learning neural network for processing, and the corresponding classification result is obtained.

[0015] Thirdly, a computer-readable storage medium is provided, including a computer program or instructions that, when executed on a computer, cause the computer to perform the following processes: Obtain resting-state functional magnetic resonance imaging (rs-fMRI) images of the brains of the subjects to be treated; The rs-fMRI images of the subjects to be treated were processed to obtain the brain functional connectivity (BFC) diagram structure of the subjects. ,in, , , For the subjects, express The set of nodes in the middle, express Adjacency matrix, scale label Nodes representing a specific spatial scale The corresponding Region of Interest (ROI) is defined for each edge. , express The corresponding heterogeneous types include: same-scale edges, cross-scale overlapping edges, and cross-scale non-overlapping edges; and, The BFC diagram structure of the subject to be processed The input is fed into a trained graph learning neural network for processing, and the corresponding classification result is obtained.

[0016] Fourthly, a computer program product is provided, comprising: a computer program or instructions that, when executed on a computer, cause the computer to perform the following processes: Obtain resting-state functional magnetic resonance imaging (rs-fMRI) images of the brains of the subjects to be treated; The rs-fMRI images of the subjects to be treated were processed to obtain the brain functional connectivity (BFC) diagram structure of the subjects. ,in, , , For the subjects, express The set of nodes in the middle, express Adjacency matrix, scale label Nodes representing a specific spatial scale The corresponding Region of Interest (ROI) is defined for each edge. , express The corresponding heterogeneous types include: same-scale edges, cross-scale overlapping edges, and cross-scale non-overlapping edges; and, The BFC diagram structure of the subject to be processed The input is fed into a trained graph learning neural network for processing, and the corresponding classification result is obtained. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a method for classifying cognitive-related brain diseases provided in an embodiment of this application; Figure 2 This is a schematic diagram of the graph learning module provided in an embodiment of this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0020] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0022] Before introducing the embodiments of this application, a brief description of the current stage of technical research of this application will be given: Cognitive-related brain disorders refer to a range of neurological diseases that affect human cognition, typically manifesting as a combination of cognitive and behavioral impairments. For example, Alzheimer's disease is a progressive neurodegenerative disease characterized by cognitive decline, memory impairment, executive dysfunction, and ultimately dementia. Autism spectrum disorder (ASD) is a neurodevelopmental disorder characterized by social and communication impairments, stereotyped ritualistic interests, and repetitive motor or verbal behaviors. Currently, cognitive-related brain disorders are considered a serious global public health challenge. Since early intervention is likely the best long-term treatment strategy for most cognitive-related brain disorders, accurate early diagnosis and classification are crucial for developing effective treatment plans for patients.

[0023] Recent studies on brain connectomics and brain dynamics have shown that the human brain can be represented by a network / graph, where nodes correspond to regions of interest (ROIs) defined by brain segmentation templates, and edges quantify the functional or structural connections between these ROIs. These connections have a unique architecture at the whole-brain scale, capable of depicting patterns of brain activity. In response, many researchers utilize medical imaging tools such as magnetic resonance imaging (MRI) to discover biomarkers associated with specific cognitive-related brain diseases, characterized by abnormal changes in brain network function and structural connectivity. Furthermore, because the structural changes in most early-stage cognitive-related brain diseases are very subtle and difficult to observe on MRI, current research widely uses resting-state functional magnetic resonance imaging (fMRI) to capture early functional abnormalities. This non-invasive brain imaging technique measures low-frequency spontaneous fluctuations in blood oxygen level-dependent (BOLD) signals. By measuring the temporally synchronized BOLD signals between different ROIs, brain functional connectivity (BFC) maps can be constructed.

[0024] Thanks to the rapid development of deep learning algorithms, especially graph neural networks, recent methods have focused on using this data-driven technique to reveal meaningful patterns in brain networks for the classification and diagnosis of cognitive-related brain diseases. Generally, the classification and diagnosis process can be formalized as a BFC (Brain Function Frame) graph classification task. First, the voxel-level BOLD (Brain Function Frame) signals of the entire brain region for each subject are converted into ROI-based BFC graphs based on specific brain segmentation templates. Then, graph representation learning based on GNNs is performed on the group-level BFC graphs to model nonlinear brain functional connectivity patterns. Finally, based on the subject's brain BFC graph representation, a classifier is used to categorize the subject into a normal control group or different stages of cognitive-related brain diseases. To mitigate spatial representation bias caused by constructing BFC graphs based on a single segmentation template, several recent methods propose employing large-scale BFC graph learning based on multiple brain templates. This allows for the comprehensive utilization of complementary brain network topological information from multiple BFC graphs for more accurate classification of cognitive-related brain diseases.

[0025] Existing methods based on large-scale multi-BFC maps have three shortcomings: (1) Existing methods mostly use completely different brain segmentation templates to divide brain regions, and these templates have almost no spatial correlation (such as the commonly used AAL template, Harvard-Oxford template, CC200 template, etc.). However, recent research in the field of neuroscience shows that the network structure of the human brain is essentially hierarchically organized, which further enables brain regions to interact in a variety of ways at multiple spatial scales during cognitive processing. In addition, the BFC networks of patients with cognitive-related brain diseases at different stages may change to varying degrees in the topological characteristics of this hierarchical modular network within a specific spatial scale or across multiple scales. Existing methods and techniques cannot model this hierarchical and multi-scale brain functional connectivity structure of the human brain; (2) Existing studies usually assume that the connections between brain regions defined at different map scales are homogeneous. The few multi-scale BFC map learning methods proposed recently only focus on the heterogeneity of connections within the same scale and between spatially overlapping cross-scale brain regions. However, existing neuroscience research evidence suggests that cognitive-related brain diseases disrupt the modularity of normal brain function networks, and patients exhibit weaker dissociation of brain functional connectivity networks than healthy controls. In other words, the functional patterns implied by connections between non-spatial-overlapping cross-scale brain regions also provide valuable information for characterizing different stages of Alzheimer's disease; (3) The representation fusion methods used in existing approaches mainly rely on late fusion strategies, such as averaging or splicing features from multiple BFC maps before classification. This intuitive late fusion strategy cannot fully capture the complex spatial dependencies between different brain maps in the early stages of cognitive-related brain diseases.

[0026] The aforementioned problems with current methods and technologies result in insufficient accuracy for existing MRI-based classification and diagnosis tasks of cognitive-related brain diseases, failing to meet clinical needs.

[0027] In response, this embodiment proposes a diagnostic classification method for cognitive-related brain diseases based on a large-scale heterogeneous brain functional connectivity network. Based on a novel end-to-end graph neural network, it uses an early fusion strategy to obtain diverse features of hierarchical heterogeneous BFC graphs derived from multi-scale brain templates. This can better understand the brain functional connectivity patterns at different stages of cognitive-related brain diseases and achieve more accurate classification of cognitive-related brain diseases.

[0028] The method proposed in this embodiment can effectively improve the accuracy and systematicness of clinical classification and diagnosis of cognitive-related brain diseases, provide clinicians with some targeted auxiliary diagnostic opinions, and thus provide patients with timely and effective early intervention, reducing the risk of the deterioration of cognitive-related brain diseases.

[0029] This embodiment proposes a solution and processing framework for the classification of cognitive-related brain diseases, as follows: Figure 1 As shown, it mainly includes the following three parts: 1. Data preprocessing and hierarchical BFC diagram construction (1) For the input subject rs-fMRI data, perform time point correction, spatial smoothing, head motion correction, noise removal and spatial normalization preprocessing operations, and perform bandpass filtering on the 0.01~0.1 Hz BOLD time domain signal to obtain the low frequency fluctuation of rs-fMRI. (2) Based on the subjects' BOLD signal sequences, this embodiment uses a set of maps provided by the Schaefer hierarchical brain template set to construct a multi-scale hierarchical BFC map for each subject. The Schaefer brain template set includes 10 brain partition templates with different spatial granularities, ranging from coarse (100 ROI) to fine (1000 ROI), each template corresponding to a different spatial scale. This embodiment uses each specific spatial scale from the Schaefer brain template set... The preprocessed BOLD signal sequence is partitioned, where , This indicates the number of different spatial scales in the Schaefer map set.

[0030] (3) For each subject, calculate the average BOLD signal of voxels in each ROI to obtain all selected scales. The region-averaged BOLD time series signal.

[0031] (4) The average signals of all ROIs in the array are concatenated to form a dimension of... Cross-scale time series matrices (such as Figure 1 (a) is shown), where Is The number of all ROIs defined in the document The number of voxels.

[0032] (5) Finally, for any subject Any two ROIs and The strength of the functional connections between them, i.e., the weights of the edges in the BFC graph, is represented by the Pearson correlation coefficient of their average time series signals: in and They represent and Time series, and They represent respectively and The time series average. Then, using... The nearest neighbor (KNN) algorithm is used to ensure that each node only associates with its nearest neighbor. Connect the most relevant nodes.

[0033] The above processing steps ultimately generated a set of BFC diagrams for all participants. ,in Indicates the subject BFC diagram express The set of nodes in the region (i.e., the ROI). express adjacency matrix This indicates the number of participants. By overlaying all brain templates from the Schaefer brain template set, a large-scale brain functional connectivity network with 5,500 nodes can be constructed for each participant.

[0034] 2. Construction of Heterogeneous BFC Graph First, the multi-scale BFC graph set obtained from the above processing is... The definition is expanded to ,in , and use scale labels Nodes in a brain template representing a specific spatial scale The corresponding ROI definition. For each edge , express Corresponding heterogeneous types, and have The specific definition is as follows: (1) Edges of the same scale For the graph Any two nodes in and If and only if and Coming from the same spatial scale When, its corresponding edge Defined as edges of the same scale and having .

[0035] (2) Cross-scale overlapping edges Given that they originate from two different map scales and Two sets of ROI and ( ), assuming Compare The spatial granularity is finer, and more ROIs are defined, that is... So, for ROI of any one of them voxel set and Any ROI A collection of voxels, and Spatial overlap between Defined as: in This represents the number of voxels within a specific ROI. Then, if the spatial scale is... ROI and spatial scale ROI The spatial overlap between them is greater than a certain threshold ,Right now Then the edge Defined as and The cross-scale overlapping hierarchical edges and have ,in As an adjustment parameter, the optimal value can be obtained through cross-experimentation.

[0036] (3) Non-overlapping edges across scales Similarly, given that they originate from two different map scales and Two sets of ROI and And there are ,for ROI of any one of them voxel set and Any ROI A set of voxels if and only if At that time, then the side Defined as and Non-overlapping edges between scales and have .

[0037] 3. Visual learning The graph learning module is the core of the algorithm framework proposed in this embodiment, aiming to explicitly learn the discriminative representation of the constructed hierarchical heterogeneous BFC graph in a fine-grained to coarse-grained manner. For example... Figure 2 As shown, for any node Its node features are obtained based on a hierarchical multi-scale learning method with spatial granularity ranging from fine to coarse. In each iteration of learning, the target node... Features are aggregated in The information of all neighboring nodes is updated, and each neighboring node is updated with information from all neighboring nodes. The contribution of feature updates depends on the information obtained in Section 2. Heterogeneous types that are connected to its neighboring nodes.

[0038] Figure 2 The process mainly consists of three parts: (a) updating node features at multiple spatial scales. For simplicity, the figure shows three scales, with the number of nodes satisfying... (b) Update the embeddings of all nodes at each scale in the BFC graph to obtain the node feature map at that specific scale. The graph is divided into scales. Next node For example, the update (marked in red). Represents subgraphs of the same scale, Represents overlapping subgraphs across scales. (c) Represent non-overlapping subgraphs across scales; generate the final BFC graph representation based on each subgraph using its corresponding aggregation function. .

[0039] 3.1 Heterogeneous Subgraph Learning for any node This section focuses on the target node in one training iteration. Taking the feature update processing flow as an example, the design details of the heterogeneous graph learning module in this embodiment are explained. Specifically, this embodiment is based on... Define the heterogeneous type of the edge to be related to Three heterogeneous subgraphs with direct edge connections are presented, and different graph learning strategies are designed for each: (1) Same-scale subgraph learning Same-scale subgraph Defined as a set containing edges of the same scale and A graph of connected neighboring nodes, that is, for middle Each neighbor node ,have and .for This embodiment uses a classic graph convolutional neural network (GraphConv) to learn... The feature maps of submaps of the same scale can be processed as follows: in express Adjacency matrix of subgraphs of the same scale The self-loop matrix, where I is the identity matrix and the diagonal matrix. The elements in are matrices The degree of each node in the middle. Indicates the first Parameters of the layer network This represents the nonlinear activation function of the network (such as the ReLU function). Indicates the first The output of the layer and the first The input of the layer.

[0040] (2) Learning of overlapping subgraphs across scales Cross-scale overlapping subgraphs To achieve cross-scale overlapping edges and A graph consisting of connected neighboring nodes, that is, for middle Each neighbor node ,have , , Given the hierarchical organization of brain functional connectivity networks across multiple topological scales, this embodiment employs a hierarchical pooling mechanism to aggregate... Each scale Node characteristics on: in, Representing scale The image below The set of nodes in This indicates a summation pooling operation.

[0041] (3) Cross-scale non-overlapping subgraph learning Cross-scale non-overlapping subgraphs To pass through non-overlapping edges across scales and A graph consisting of connected neighboring nodes, that is, for middle Each neighbor node ,have , , . and in and The learning process is different from that of training. Message passing on the platform is done independently. At various spatial scales The following is feature aggregation based on the heterogeneous graph convolution operation (HGraphConv): in, Representing scale The image below Node set in It is used to obtain The aggregation function embedding the features of neighboring nodes. It can be implemented based on average computation, maximum computation, or attention mechanisms.

[0042] 3.2 Subgraph Feature Aggregation (1) In the case of Cross-scale subplots at various spatial scales ( and After performing the above learning steps, this embodiment integrates the two subgraphs based on a multilayer perceptron (MLP) layer. The stitching features of each cross-scale overlapping subgraph and non-overlapping subgraph are used to generate... Cross-scale feature embedding: in, From the scale Down Embedded features extracted from cross-scale overlapping subgraphs, Indicates from scale Down Features obtained from cross-scale non-overlapping subgraphs. This represents a join function consisting of two join operators. The process is shown in Figure 2.

[0043] (2) Aggregate cross-scale feature embeddings with same-scale feature embeddings, and generate spatial scale based on the following formula. The following node features: in, Representing scale The next node Same-scale and cross-scale feature embeddings ( (The output dimension of the convolutional network used). and These are the corresponding learnable aggregated weight vectors. Furthermore, this embodiment learns these two weight vectors based on an adaptive strategy: for each scale... This embodiment utilizes a learnable set of weight vectors. The softmax function on the above is obtained and ,in And in each network layer there is .

[0044] 3.3 Full-image feature fusion Since the contribution of node features at different spatial scales to the final BFC graph representation varies, this embodiment uses a multi-head self-attention mechanism to fuse the node feature maps at each spatial scale obtained in Section 3.2.

[0045] (1) For each spatial scale The Readout function is used to aggregate the embedding features of all nodes at this spatial scale to obtain a graph representation at a specific spatial scale. ( (The output dimension of the convolutional network used).

[0046] (2) Concatenate the graphical representations at all spatial scales to obtain .

[0047] (3) Based on the results obtained in the previous step Generate a query matrix A key matrix A value matrix ,in , , It is a learnable linear mapping matrix.

[0048] (4) Obtain the final representation of the BFC graph based on multi-head self-attention. : in, This represents the number of attention heads; multiple attention heads are concatenated along the embedding dimension and then weighted using a weight matrix. Mapped to the original dimension.

[0049] 4. Model Training During the model training phase, a multiple cross-training method can be used to divide the training dataset for multiple training iterations. This method uses cross-entropy loss to guide parameter updates during model training. in Indicates the number of samples, and These represent the true label and the predicted label, respectively.

[0050] It should be noted that: Step (5) of “1. Data Preprocessing and Hierarchical BFC Graph Construction”: Finally, for any subject Any two ROIs and The functional connectivity strength between them, i.e., the edge weights in the BFC graph, is represented by the Pearson correlation coefficient of their average time series signals. This step involves the participants... Two ROIs and The functional connectivity strength between them can be calculated not only based on Pearson's correlation coefficient of their average time series signals, but also through Spearman correlation coefficient, partial correlation function, cross-correlation function, and other correlation calculation methods.

[0051] This application also provides a cognitive-related brain disease classification device, including: The acquisition unit is used to acquire resting-state functional magnetic resonance imaging (rs-fMRI) images of the brain of the subjects to be processed. The initial processing unit is used to process the rs-fMRI images of the subject to be processed to obtain the brain functional connectivity (BFC) diagram structure of the subject. ,in, , , For the subjects, express The set of nodes in the middle, express Adjacency matrix, scale label Nodes representing a specific spatial scale The corresponding Region of Interest (ROI) is defined for each edge. , express The corresponding heterogeneous types include: same-scale edges, cross-scale overlapping edges, and cross-scale non-overlapping edges; and, The classification unit is used to classify the BFC diagram structure of the subject to be processed. The input is fed into a trained graph learning neural network for processing, and the corresponding classification result is obtained.

[0052] In one possible design, the initial processing unit specifically includes: The preprocessing unit is used to preprocess the rs-fMRI images of the subject to be processed to obtain the brain oxygenation level-dependent BOLD signal sequence of the subject to be processed; and, The construction unit is used to construct a multi-scale hierarchical BFC map of the subject's brain using brain partition templates at different spatial scales; calculate the average value of the subject's BOLD signal within voxels of ROIs in the multi-scale hierarchical BFC map; concatenate the average values ​​of the subject's BOLD signals to form a cross-scale time series matrix; and construct the initial BFC map structure of the subject based on the cross-scale time series matrix. ,in, Indicates the subject BFC diagram, express The set of nodes in the middle, express The adjacency matrix, The number of subjects is indicated; and the initial BFC graph structure of the subjects to be processed is expanded to the BFC graph structure of the subjects to be processed.

[0053] In one possible design approach, the weights of the edges of the initial BFC graph structure of the subject to be processed are represented by the Pearson correlation coefficient, Spearman correlation coefficient, partial correlation function, or cross-correlation function of the average time series signal of the edges.

[0054] In one possible design approach, the preprocessing includes: time point correction, spatial smoothing, head motion correction, noise removal, spatial normalization, and / or bandpass filtering.

[0055] In one possible design approach, the same-scale edge is defined as: for a graph Any two nodes in and If and only if and Coming from the same spatial scale When, its corresponding edge Defined as edges of the same scale and having ; The cross-scale overlapping edge is defined as follows: given two different map scales... Two sets of ROI and Assuming Compare With finer spatial granularity and more ROIs defined, then for ROI of any one of them voxel set and Any ROI A collection of voxels, and Spatial overlap between Defined as: in, This represents the number of voxels within a specific ROI, if Then the edge Defined as and The cross-scale overlapping hierarchical edges and have , To adjust the parameters; The cross-scale non-overlapping edge is defined as: given two different map scales... Two sets of ROI and ,for ROI of any one of them voxel set and Any ROI A set of voxels if and only if At that time, then the side Defined as and Non-overlapping edges between scales and have .

[0056] In one possible design approach, the classification unit specifically includes: The heterogeneous subgraph feature aggregation unit is used to process, for any node in the BFC graph structure of the subject to be processed, the subject's same-scale subgraph corresponding to the same-scale edge using a graph convolutional neural network to obtain the subject's same-scale feature embedding; to aggregate the subject's cross-scale overlapping subgraph corresponding to the cross-scale overlapping edge using a hierarchical pooling mechanism to obtain the subject's cross-scale overlapping feature embedding; to aggregate the subject's cross-scale non-overlapping subgraph corresponding to the cross-scale non-overlapping edge using heterogeneous graph convolution operations to obtain the subject's cross-scale non-overlapping feature embedding; to concatenate the subject's cross-scale overlapping feature embedding and the subject's cross-scale non-overlapping feature embedding to obtain the subject's cross-scale feature embedding; and to aggregate the subject's same-scale feature embedding and the subject's cross-scale feature embedding to obtain the subject's node embedding features. A full-image feature fusion unit is configured to perform full-scale fusion of the embedded features of all nodes in the BFC graph structure of the subject to be processed, thereby obtaining the full-image fused features of the subject to be processed; and... The classification subunit is used to process the full-image fusion features of the subject to be processed to obtain the classification result.

[0057] In one possible design approach, the classification unit employs the cross-entropy loss function Lce: in, Indicates the number of samples, and These represent the true label and the predicted label, respectively.

[0058] This application also provides an electronic device, including: a processor, and a memory coupled to the processor, the memory being used to store a computer program; the processor being used to execute the computer program stored in the memory, so that the electronic device performs the method as described in any of the above embodiments.

[0059] Electronic devices can be computing devices such as desktop computers, laptops, handheld computers, and cloud servers. These electronic devices may include, but are not limited to, processors and memory.

[0060] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting various parts of the device via various interfaces and lines.

[0061] The memory can be used to store the computer program, and the processor implements various functions of the electronic device by running or executing the computer program stored in the memory and calling the data stored in the memory.

[0062] The memory may primarily include a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function, etc.; the data storage area may store data created based on the use of the mobile phone, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0063] This application also provides a storage medium, which is a computer-readable storage medium. The computer program is stored in the computer-readable storage medium, and when executed by a processor, the computer program can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0064] This application also provides a computer program product, including: a computer program or instructions that, when the computer program or instructions are run on a computer, cause the computer to perform any of the above possible implementation methods.

[0065] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications are also considered to be within the scope of protection of this application.

Claims

1. A device for classifying cognitive-related brain diseases, characterized in that, include: The acquisition unit is used to acquire resting-state functional magnetic resonance imaging (rs-fMRI) images of the brain of the subjects to be processed. The initial processing unit is used to process the rs-fMRI images of the subject to be processed to obtain the brain functional connectivity (BFC) diagram structure of the subject. ,in, , , For the subjects, express The set of nodes in the middle, express Adjacency matrix, scale label Nodes representing a specific spatial scale The corresponding Region of Interest (ROI) is defined for each edge. , express The corresponding heterogeneous types include: same-scale edges, cross-scale overlapping edges, and cross-scale non-overlapping edges; and, The classification unit is used to classify the BFC diagram structure of the subject to be processed. The input is fed into a trained graph learning neural network for processing, and the corresponding classification result is obtained.

2. The cognitive-related brain disease classification device as described in claim 1, characterized in that, The initial processing unit specifically includes: The preprocessing unit is used to preprocess the rs-fMRI images of the subject to be processed to obtain the brain oxygenation level-dependent BOLD signal sequence of the subject to be processed; and, The construction unit is used to construct a multi-scale hierarchical BFC map of the subject's brain using brain partition templates at different spatial scales; calculate the average value of the subject's BOLD signal within voxels of ROIs in the multi-scale hierarchical BFC map; concatenate the average values ​​of the subject's BOLD signals to form a cross-scale time series matrix; and construct the initial BFC map structure of the subject based on the cross-scale time series matrix. ,in, Indicates the subject BFC diagram, express The set of nodes in the middle, express The adjacency matrix, The number of subjects is indicated; and the initial BFC graph structure of the subjects to be processed is expanded to the BFC graph structure of the subjects to be processed.

3. The cognitive-related brain disease classification device as described in claim 2, characterized in that, The weights of the edges in the initial BFC graph structure of the subjects to be processed are represented by the Pearson correlation coefficient, Spearman correlation coefficient, partial correlation function, or cross-correlation function of the average time series signal of the edges.

4. The cognitive-related brain disease classification device as described in claim 2, characterized in that, The preprocessing includes: time point correction, spatial smoothing, head motion correction, noise removal, spatial normalization and / or bandpass filtering.

5. The cognitive-related brain disease classification device as described in claim 1, characterized in that, The same-scale edge is defined as: for a graph Any two nodes in and If and only if and Coming from the same spatial scale When, its corresponding edge Defined as edges of the same scale and having ; The cross-scale overlapping edge is defined as follows: given two different map scales... Two sets of ROI and Assuming Compare With finer spatial granularity and more ROIs defined, then for ROI of any one of them voxel set and Any ROI A collection of voxels and Spatial overlap between Defined as: in, Indicates the number of voxels within a specific ROI, if Then the edge Defined as and The cross-scale overlapping hierarchical edges between and have , To adjust the parameters; The cross-scale non-overlapping edge is defined as: given two different map scales... Two sets of ROI and ,for ROI of any one of them voxel set and Any ROI A set of voxels if and only if At that time, then the side Defined as and Non-overlapping edges between scales and have .

6. The cognitive-related brain disease classification device as described in claim 1, characterized in that, The classification unit specifically includes: The heterogeneous subgraph feature aggregation unit is used to process, for any node in the BFC graph structure of the subject to be processed, the subject's same-scale subgraph corresponding to the same-scale edge using a graph convolutional neural network to obtain the subject's same-scale feature embedding; to aggregate the subject's cross-scale overlapping subgraph corresponding to the cross-scale overlapping edge using a hierarchical pooling mechanism to obtain the subject's cross-scale overlapping feature embedding; to aggregate the subject's cross-scale non-overlapping subgraph corresponding to the cross-scale non-overlapping edge using heterogeneous graph convolution operations to obtain the subject's cross-scale non-overlapping feature embedding; to concatenate the subject's cross-scale overlapping feature embedding and the subject's cross-scale non-overlapping feature embedding to obtain the subject's cross-scale feature embedding; and to aggregate the subject's same-scale feature embedding and the subject's cross-scale feature embedding to obtain the subject's node embedding features. A full-image feature fusion unit is configured to perform full-scale fusion of the embedded features of all nodes in the BFC graph structure of the subject to be processed, thereby obtaining the full-image fused features of the subject to be processed; and... The classification subunit is used to process the full-image fusion features of the subject to be processed to obtain the classification result.

7. The cognitive-related brain disease classification device as described in claim 1, characterized in that, The classification unit uses the cross-entropy loss function Lce: in, Indicates the number of samples, and These represent the true label and the predicted label, respectively.

8. An electronic device, characterized in that, The electronic device includes: a processor, and a memory coupled to the processor. The memory is used to store computer programs; The processor is configured to execute the computer program stored in the memory, causing the electronic device to perform the following process: Obtain resting-state functional magnetic resonance imaging (rs-fMRI) images of the brains of the subjects to be treated; The rs-fMRI images of the subjects were processed to obtain the brain functional connectivity (BFC) diagram structure of the subjects. ,in, , , For the subjects, express The set of nodes in the middle, express Adjacency matrix, scale label Nodes representing a specific spatial scale The corresponding Region of Interest (ROI) is defined for each edge. , express The corresponding heterogeneous types include: same-scale edges, cross-scale overlapping edges, and cross-scale non-overlapping edges; and, The BFC diagram structure of the subject to be processed The input is fed into a trained graph learning neural network for processing, and the corresponding classification result is obtained.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a computer program or instructions that, when executed on a computer, cause the computer to perform the following processes: Obtain resting-state functional magnetic resonance imaging (rs-fMRI) images of the brains of the subjects to be treated; The rs-fMRI images of the subjects were processed to obtain the brain functional connectivity (BFC) diagram structure of the subjects. ,in, , , For the subjects, express The set of nodes in the middle, express Adjacency matrix, scale label Nodes representing a specific spatial scale The corresponding Region of Interest (ROI) is defined for each edge. , express The corresponding heterogeneous types include: same-scale edges, cross-scale overlapping edges, and cross-scale non-overlapping edges; and, The BFC diagram structure of the subject to be processed The input is fed into a trained graph learning neural network for processing, and the corresponding classification result is obtained.

10. A computer program product, characterized in that, The computer program product includes: a computer program or instructions that, when executed on a computer, cause the computer to perform the following processes: Obtain resting-state functional magnetic resonance imaging (rs-fMRI) images of the brains of the subjects to be treated; The rs-fMRI images of the subjects were processed to obtain the brain functional connectivity (BFC) diagram structure of the subjects. ,in, , , For the subjects, express The set of nodes in the middle, express Adjacency matrix, scale label Nodes representing a specific spatial scale The corresponding Region of Interest (ROI) is defined for each edge. , express The corresponding heterogeneous types include: same-scale edges, cross-scale overlapping edges, and cross-scale non-overlapping edges; and, The BFC diagram structure of the subject to be processed The input is fed into a trained graph learning neural network for processing, and the corresponding classification result is obtained.

Images