An autism federated collaborative magnetic resonance imaging data classification method based on graph topology prototype
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- UNIV OF ELECTRONICS SCI & TECH OF CHINA
- Filing Date
- 2026-03-04
- Publication Date
- 2026-06-16
Smart Images
Figure CN122223402A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of medical image analysis and privacy computing technology, and specifically relates to a distributed brain image collaborative diagnosis method based on topological prototype enhancement for multicenter functional magnetic resonance imaging (fMRI) data. Background Technology
[0002] In diagnostic research for conditions such as autism, brain functional network analysis based on functional magnetic resonance imaging (fMRI) plays a crucial role. However, it faces two major challenges in practical clinical applications and scientific research: First, multi-center data exhibits significant heterogeneity due to differences in acquisition equipment, scanning parameters, and subject groups, resulting in poor generalization ability of models trained at a single center. Second, due to the privacy sensitivity of medical data and legal and regulatory restrictions, direct sharing of raw data across institutions is often not feasible. Currently, technologies addressing these issues mainly include: 1. Traditional federated learning methods: such as FedAvg (see "B. McMahan et al., Communication-Efficient Learning of Deep Networks from Decentralized Data, AISTATS, 2017"), which collaboratively trains models by aggregating model parameters. However, under highly heterogeneous data, the averaging of global model parameters leads to feature drift and severe performance degradation. 2. Prototype-based federated learning methods: such as FedProto (see "Y. Tan et al., FedProto: Federated Prototype Learning across Heterogeneous Clients, AAAI, 2022"). These methods protect privacy and mitigate heterogeneity by exchanging class prototypes rather than model parameters. However, existing prototyping methods are mostly designed for natural images, ignoring the complex topological semantics of brain functional networks, and suffer from problems such as the lack of discriminative power in the extracted prototypes. Summary of the Invention
[0003] The purpose of this invention is to overcome the shortcomings of the prior art by proposing a distributed brain imaging collaborative diagnosis method based on topological prototype enhancement. By combining the topological representation capabilities of graph neural networks with the prototype learning mechanism, efficient collaboration and accurate diagnosis of multi-center brain imaging data can be achieved without sharing the original data.
[0004] To achieve the above objectives, the technical solution of the present invention is a graph topology prototype-based federated collaborative magnetic resonance imaging data classification method for autism, which includes the following steps:
[0005] S1. Obtain resting-state functional magnetic resonance imaging data from multiple independent centers, and perform time correction, head movement correction, spatial standardization and denoising operations on the data; then, extract brain region time series for each subject based on predefined brain region templates, calculate the functional connectivity strength between brain regions, construct the corresponding functional connectivity matrix, and form a brain functional map G with brain regions as nodes and functional connectivity as edges, which serves as the input for subsequent model training.
[0006] S2. In the local client, an attention modeling mechanism based on adaptive adjustment of channel weight and structural weight is introduced into the constructed brain function map to weight the importance of different functional connections and their topological structures, and enhance the expression of key connection patterns related to category discrimination; on this basis, graph-level features are aggregated to generate graph topological prototype representations of the corresponding categories, which are used to characterize the common features of samples of the same category in the topological space.
[0007] S3. Based on the local category graph topological prototype generated in step S2, calculate the difference vector between the target category prototype and the opposing category prototype, and generate a contrast mask through a gating mechanism; the contrast mask acts on the corresponding brain function map G in the form of residuals to enhance the topological components related to category discrimination, thereby obtaining the prototype-guided modulated graph features G. ， ;
[0008] S4. The image feature G after prototype mask modulation ， The input is a dynamic graph convolutional encoder. Through the dynamic graph convolutional layer that updates node features and edge features collaboratively, the adjacency relationship is dynamically modeled based on the attention mechanism, realizing the adaptive propagation of information in the brain functional map and obtaining a high-level graph representation that simultaneously integrates local connectivity and global topology.
[0009] S5. During federated communication, each local client executes a two-stage training strategy based on the model constructed in steps S2–S4: the first stage focuses on optimizing the consistency of the graph topology prototype and the prototype-guided feature enhancement objective; the second stage, while maintaining prototype constraints, jointly optimizes the classification loss and the prototype enhancement loss; the server only receives the graph topology prototype representation uploaded by each center, performs weighted aggregation on the graph topology prototype representation to generate a global graph topology prototype, and distributes the aggregated global prototype to each local client to guide the next round of local model training;
[0010] S6. After completing federated training, the newly acquired brain function map samples are classified using the trained local client model and global graph topology prototype: First, the resting-state functional magnetic resonance imaging data to be classified is preprocessed and functional connectivity matrix constructed in step S1 to form the corresponding brain function map G; then, map G is input into the trained dynamic graph convolutional encoder, and the graph features are modulated using the global graph topology prototype to obtain the prototype-enhanced graph representation; finally, the graph representation is input into the classifier to determine the category of the new brain function map samples, thereby completing the autism brain image recognition and classification based on topology prototype enhancement.
[0011] Furthermore, the graph topology prototype constructed in step 2 is represented as follows:
[0012] ;
[0013] in, This indicates that the extracted sample features have the same dimensions as the corresponding node and edge features. Indicates the number of samples. The exponential parameter for generalized average pooling.
[0014] Furthermore, in step 3, the comparison mask is applied to the corresponding graph feature representation in the form of residual modulation:
[0015]
[0016] in, This represents the node or edge features after adding a contrast mask. Represents the characteristics of the corresponding node or edge. To correspond to the mask features, This is the scaling factor. This represents the Hadamard product, which is the product of corresponding elements.
[0017] Furthermore, the dynamic graph neural network encoder in step 4 includes 5 NEGAN modules, wherein each NEGAN module consists of an Attention module, a GCN layer of node features or edge features, and a BatchNorm module.
[0018] Furthermore, the training method in step 5 is as follows:
[0019] The loss function used for training the calculation process from step 2 to step 4:
[0020] ;
[0021] ;
[0022] in, Let represent the loss function for the first stage of training. This represents the loss function for the second stage of training. This represents the classification loss based on graph representation, constructed from the classic cross-entropy loss. The alignment loss between the local topology prototype and the global topology prototype is defined as:
[0023] ;
[0024] This represents the graph topology prototype generated by the local client. This represents the global graph topology prototype obtained from server-side aggregation. Represents the L2 norm;
[0025] The feature enhancement loss based on topological prototypes is defined as follows:
[0026] ;
[0027] in, This represents the edge or node features corresponding to the sample; This represents a graph topological prototype corresponding to the true category of the sample. and These represent the graph topological prototypes of the two types of samples, respectively. Temperature coefficient;
[0028] The cosine similarity between vectors a and b is calculated as follows:
[0029] ;
[0030] The two parameters used to describe the sim() calculation method have no practical significance. During the training process, the Adam optimization method is used to update the network parameters. After completing multiple rounds of local training and prototype aggregation, the final model for multi-center brain imaging collaborative diagnosis is obtained.
[0031] The beneficial effects of this invention are as follows: This invention proposes a prototype-based distributed collaborative training mechanism for brain network topological semantics. By introducing a federated collaborative learning mechanism based on graph topological prototypes, it achieves collaborative training of multi-center brain functional network models without sharing the original functional magnetic resonance imaging (fMRI) data, effectively alleviating the problem of model performance degradation under conditions of restricted medical data privacy. Simultaneously, by constructing class-level graph topological prototypes and enhancing feature representations through prototype guidance during local training, the model's ability to distinguish different brain functional states is improved, reducing the impact of distribution offset between different acquisition centers on model performance. Furthermore, by adopting a communication strategy that aggregates topological prototypes only on the server side, the communication overhead during federated training is significantly reduced, while avoiding the privacy leakage risk caused by direct sharing of model parameters. Attached Figure Description
[0032] Figure 1 This is a schematic diagram of the model flow of a graph topology-based federalized collaborative magnetic resonance imaging data classification method for autism based on the present invention.
[0033] Figure 2 This is the prototype boot mask obtained using the method of the present invention in the embodiment.
[0034] Figure 3 This is a graph showing the change in model accuracy with training epochs when conducting experiments on the ABIDE I multicenter dataset in an embodiment of the present invention. Detailed Implementation
[0035] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.
[0036] S1. Download the ABIDE data as a training dataset; perform a unified preprocessing operation on the functional magnetic resonance imaging (fMRI) data on each local client, preferably using the fmriprep preprocessing workflow; then, using the AAL template, extract brain region time series for each subject, and calculate the Pearson correlation coefficient between any two brain region time series to construct a functional connectivity matrix between brain regions as edge features, with a size of 116×116. Optimize the degree, betweenness centrality, and global efficiency of the functional connectivity matrix to construct node features of size 116×3, and use the brain functional map as input data for the subsequent graph neural network model.
[0037] S2. Design a graph topology feature extraction network, specifically a one-layer SE module and a two-layer Attention module implemented based on the open-source framework PyTorch. Extract the topological features of the brain function maps constructed in step S1 on the local client, and then aggregate multiple brain function map features under the same category using generalized average pooling to construct the graph topology prototype representation for the corresponding category:
[0038] ;
[0039] in This indicates that the extracted sample features have the same dimensions as the corresponding node and edge features. Indicates the number of samples. The exponential parameter for generalized average pooling is preferably 2.5.
[0040] S3. Based on the local category graph topology prototype constructed in step S2, calculate the topological difference between the target category prototype and the non-target category prototype, and generate a prototype guiding mask with the same node feature and edge feature dimensions using the Sigmoid activation function; apply the mask to the corresponding graph feature representation in the form of residual modulation:
[0041] ;
[0042] in Represents the characteristics of the corresponding node or edge. To correspond to the mask features, This is the scaling factor, initially set to 1.5. The prototype bootstrap mask is as follows: Figure 2 As shown.
[0043] S4. Construct a dynamic graph neural network encoder, consisting of 5 classic NEGAN modules. Each NEGAN module comprises an Attention module, and node features and edge features each correspond to a GCN and a BatchNorm module. Input the prototype augmented graph features obtained in step S3 into the dynamic graph convolutional encoder as input to the final classification module.
[0044] S5. During federated training, the client employs a two-stage training strategy to optimize the model, while the server only receives and aggregates the graph topology prototypes uploaded by each client. The total global training rounds T are set to 20, and the local training rounds are set to 5. The loss function for model training is:
[0045] ;
[0046] ;
[0047] in, This represents the classification loss based on graph representation, constructed from the classic cross-entropy loss. The alignment loss between the local topology prototype and the global topology prototype is defined as:
[0048] ;
[0049] This represents the graph topology prototype generated by the local client. This represents the global graph topology prototype obtained from server-side aggregation. This represents the L2 norm.
[0050] The feature enhancement loss based on topological prototypes is defined as follows:
[0051] ;
[0052] in This represents the edge or node features corresponding to the sample; This represents a graph topological prototype corresponding to the true category of the sample. and These represent the graph topological prototypes of the two types of samples, respectively. The temperature coefficient is preferably 0.02. The cosine similarity between vectors a and b is calculated as follows:
[0053] ;
[0054] During training, the Adam optimization method is used to update the network parameters. After multiple rounds of local training and prototype aggregation, the final model for multi-center brain imaging collaborative diagnosis is obtained.
[0055] During the testing phase, such as Figure 1 The model inputs the functional magnetic resonance imaging (fMRI) data of the test subjects into the trained model, and the model outputs the corresponding diagnostic results. Experimental results show that the method of this invention can achieve effective collaborative modeling of multi-center brain imaging data without sharing the original data, and obtain stable diagnostic performance.
[0056] Those skilled in the art should understand that the above embodiments are only used to illustrate the technical solutions of the present invention and do not constitute a limitation on the scope of protection of the present invention; various modifications or substitutions made thereto without departing from the spirit and substance of the present invention should fall within the scope of protection of the present invention.
[0057] Table 1 shows the classification accuracy comparison results on the ABIDE I dataset in the embodiments of the present invention, including the performance of the complete model and the model after removing different functional modules.
[0058]
Claims
1. A graph topology prototype-based method for classifying autism-related federated collaborative magnetic resonance imaging data, comprising the following steps: S1. Obtain resting-state functional magnetic resonance imaging data from multiple independent centers, and perform time correction, head movement correction, spatial standardization and denoising operations on the data; then, extract brain region time series for each subject based on predefined brain region templates, calculate the functional connectivity strength between brain regions, construct the corresponding functional connectivity matrix, and form a brain functional map G with brain regions as nodes and functional connectivity as edges, which serves as the input for subsequent model training. S2. In the local client, an attention modeling mechanism based on adaptive adjustment of channel weight and structural weight is introduced into the constructed brain function map to weight the importance of different functional connections and their topological structures, thereby enhancing the expression of key connection patterns related to category discrimination. Based on this, the graph-level features are aggregated to generate graph topological prototype representations of the corresponding categories, which are used to characterize the common features of samples of the same category in the topological space. S3. Based on the local category graph topological prototype generated in step S2, calculate the difference vector between the target category prototype and the opposing category prototype, and generate a contrast mask through a gating mechanism; the contrast mask acts on the corresponding brain function map G in the form of residuals to enhance the topological components related to category discrimination, thereby obtaining the prototype-guided modulated graph features G. ， ; S4. The image feature G after prototype mask modulation ， The input is a dynamic graph convolutional encoder. Through the dynamic graph convolutional layer that updates node features and edge features collaboratively, the adjacency relationship is dynamically modeled based on the attention mechanism, realizing the adaptive propagation of information in the brain functional map and obtaining a high-level graph representation that simultaneously integrates local connectivity and global topology. S5. During federated communication, each local client executes a two-stage training strategy based on the model built in steps S2–S4: the first stage focuses on optimizing the consistency of the graph topology prototype and the prototype-guided feature enhancement objective; In the second stage, while maintaining prototype constraints, the classification loss and prototype augmentation loss are jointly optimized. The server only receives the graph topology prototype representations uploaded by each center, performs weighted aggregation on the graph topology prototype representations to generate a global graph topology prototype, and distributes the aggregated global prototype to each local client to guide the next round of local model training. S6. After completing federated training, the newly acquired brain function map samples are classified using the trained local client model and global graph topology prototype: First, the resting-state functional magnetic resonance imaging data to be classified is preprocessed and functional connectivity matrix constructed in step S1 to form the corresponding brain function map G; then, map G is input into the trained dynamic graph convolutional encoder, and the graph features are modulated using the global graph topology prototype to obtain the prototype-enhanced graph representation; finally, the graph representation is input into the classifier to determine the category of the new brain function map samples, thereby completing the autism brain image recognition and classification based on topology prototype enhancement.
2. The autism federated collaborative magnetic resonance imaging data classification method based on graph topology prototype as described in claim 1, characterized in that, The graph topology prototype constructed in step 2 is represented as follows: ； in, This indicates that the extracted sample features have the same dimensions as the corresponding node and edge features. Indicates the number of samples. The exponential parameter for generalized average pooling.
3. The autism federated collaborative magnetic resonance imaging data classification method based on graph topology prototype as described in claim 1, characterized in that, In step 3, the contrast mask is applied to the corresponding graph feature representation in the form of residual modulation: ； in, This represents the node or edge features after adding a contrast mask. Represents the characteristics of the corresponding node or edge. To correspond to the mask features, This is the scaling factor. This represents the Hadamard product, which is the product of corresponding elements.
4. The autism federated collaborative magnetic resonance imaging data classification method based on graph topology prototype as described in claim 1, characterized in that, The dynamic graph neural network encoder in step 4 includes 5 NEGAN modules, each of which consists of an Attention module, a GCN layer for node features or edge features, and a BatchNorm module.
5. The autism federated collaborative magnetic resonance imaging data classification method based on graph topology prototype as described in claim 1, characterized in that, The training method for step 5 is as follows: The loss function used for training the calculation process from step 2 to step 4: ； ； in, Let represent the loss function for the first stage of training. This represents the loss function for the second stage of training. This represents the classification loss based on graph representation, constructed from the classic cross-entropy loss. The alignment loss between the local topology prototype and the global topology prototype is defined as: ； This represents the graph topology prototype generated by the local client. This represents the global graph topology prototype obtained from server-side aggregation. Represents the L2 norm; The feature enhancement loss based on topological prototypes is defined as follows: ； in, This represents the edge or node features corresponding to the sample; This represents a graph topological prototype corresponding to the true category of the sample. and These represent the graph topological prototypes of the two types of samples, respectively. Temperature coefficient; The cosine similarity between vectors a and b is calculated as follows: ； The two parameters used to describe the sim() calculation method have no practical significance. During the training process, the Adam optimization method is used to update the network parameters. After completing multiple rounds of local training and prototype aggregation, the final model for multi-center brain imaging collaborative diagnosis is obtained.