SELF-CLUSTERING GRAPH TRANSFORMER APPROACH TO MODEL RESTING STATE FUNCTIONAL BRAIN ACTIVITY

Bishal Thapaliya, Esra Akbas, Ram Sapkota Bhaskar Ray Vince Calhoun Jingyu Liu · 2025

Read SELF-CLUSTERING GRAPH TRANSFORMER APPROACH TO MODEL RESTING STATE FUNCTIONAL BRAIN ACTIVITY by Bishal Thapaliya, Esra Akbas, Ram Sapkota Bhaskar Ray Vince Calhoun Jingyu Liu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-16345.txt.

Introduction

Bishal Thapaliya and five co-authors introduce a self-clustering graph transformer for resting-state brain activity, in work from 2025. Resting-state fMRI records brain organization without a task or stimulus. Standard graph transformers update all nodes in the same way, which can blur the brain's subnetwork structure.

Their model, called SCGT, uses functional connectivity features and updates nodes cluster by cluster. That design lets the model learn and show the subclusters it finds. Tested on the Adolescent Brain Cognitive Development dataset, with 7,957 participants, SCGT beat the standard graph transformer and other recent models at predicting cognitive score and classifying gender. The authors present it as a tool for interpreting brain networks.

Catalog record

CREATORBishal Thapaliya, Esra Akbas, Ram Sapkota Bhaskar Ray Vince Calhoun Jingyu Liu
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE20 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-16345

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

SELF-CLUSTERING GRAPH TRANSFORMER APPROACH TO MODEL RESTING STATE FUNCTIONAL BRAIN ACTIVITY Bishal Thapaliya, Esra Akbas, Ram Sapkota Bhaskar Ray Vince Calhoun Jingyu Liu Abstract Resting-state functional magnetic resonance imaging (rs-fMRI) offers valuable insights into the human brain’s functional organization and is a powerful tool for investigating the relationship between brain function and cognitive processes as it allows for the functional organization of the brain to be captured without relying on a specific task or stimuli. In this study, we introduce a novel attention mechanism for graphs with subnetworks, named Self Clustering Graph Transformer ( SCGT ), designed to handle the issue of uniform node updates in graph transformers. By using static functional connectivity (FC) correlation features as input to the transformer model, SCGT effectively captures the sub-network structure of the brain by performing cluster-specific updates to the nodes unlike uniform node updates like vanilla graph transformers, further allowing us to learn and interpret the subclusters. We validate our approach on the Adolescent Brain Cognitive Development (ABCD) dataset, comprising 7,957 participants, for the prediction of total cognitive score and gender classification. Our results demonstrate that SCGT outperforms the vanilla graph transformer method, and other recent models, offering a promising tool for modeling brain functional connectivity and interpreting the underlying subnetwork structures. I Introduction Resting-state functional magnetic resonance imaging (rs-fMRI) provides an unprecedented opportunity to delve into the human brain’s functional organization [ 1 ] . By measuring spontaneous fluctuations in the blood-oxygen-level-dependent (BOLD) signal when the brain is at rest, rs-fMRI reveals intrinsic functional connectivity patterns among different brain regions. Although traditional MRI studies in the literature focused on structural brain measures for various phenotypes ( [ 2 , 3 , 4 ] ), rapidly growing studies have emerged to investigate the prediction of intelligence based on brain functional features ( [ 5 , 6 , 7 , 8 ] ). Functional connectivity (FC) is defined as the degree of temporal correlation between regions of the brain computed using time series of BOLD signals, and rs-fMRI FC provides a comprehensive view of the brain’s intrinsic organization. Graph neural networks (GNNs) have emerged as powerful tools for modeling complex relational data, such as brain networks ( [ 9 , 10 ] ). However, most existing GNNs assume uniform node updates across the entire graph, which poses challenges when dealing with the intricate subnetwork structures of the brain connectome. Recall that GNNs update each node’s embedding using information from the node’s neighbors. The information exchange is mediated by trainable weight matrices. The same weights are uniformly applied to all nodes. However, such uniform node updates can make the GNN overlook rich local structures in the graph. Nodes within the same cluster exhibit higher similarity and stronger dependencies than nodes from different clusters. This case applies completely to Graph Transformers (GTs) as well. Current GNNs and GTs can miss these nuanced local patterns and community-specific behaviors by treating all nodes identically. Addressing this issue, BrainGNN [ 11 ] and BrainRGIN [ 12 ] introduced a clustering-based embedding technique within the graph convolutional layer, enabling nodes belonging to distinct clusters to learn embeddings in a customized manner. However, there remains a need for attention mechanisms, more powerful than graph convolutions, that can effectively handle subclustered graphs and allow for the interpretation of learned subclusters. In this study, we propose a novel attention mechanism for graphs with subnetworks, named Self Clustering Graph Transformer ( SCGT ), designed to address the issue of uniform node updates in GTs. By using static FC correlation features as input to the transformer model, our approach leverages a self-clustering graph attention block to capture the subnetwork structure of the brain. This allows us to learn and interpret the subclusters within the brain’s functional networks. Unlike traditional transformer models, which apply uniform attention mechanisms across all nodes, SCGT employs a specialized attention mechanism tailored for subclustered graphs. We validate our approach using the Adolescent Brain Cognitive Development (ABCD) dataset, comprising 7,957 participants, for the prediction of total cognitive score and gender classification. Our results demonstrate that SCGT outperforms the vanilla graph transformer method and is competitive with other baseline models, offering a promising tool for modeling brain functional connectivity and interpreting the underlying subnetwork structures. II Materials and Methods II-A The ABCD Dataset The Adolescent Brain Cognitive Development (ABCD) study is a large ongoing longitudinal study following youths from age 9-10 into late adolescence to understand factors that influence individual brain development. Participants were recruited from 21 sites across the United States to represent various demographic variables. Data used in this study were from 7,957 children aged 9–10 at baseline, including rs-fMRI images, total cognitive composite scores measured by the NIH Toolbox Cognitive Battery, and gender information. Data were split into training (n = 4773), validation (n = 1592), and test (n = 1592) subsets. Age and site effects were regressed out from the cognitive scores to ensure that only the relevant intelligence features are captured by our graph model. We conducted preprocessing on the raw rs-fMRI data utilizing a combination of FSLv6.0 and SPM12 toolboxes, encompassing several key steps, namely: 1) rigid head motion correction; 2) distortion correction; 3) removal of dummy scans; 4) normalization to the Montreal Neurological Institute space; and 5) a 6mm Gaussian kernel smoothing. Subsequently, we employed a fully automated spatially constrained independent component analysis to extract 100 intrinsic components using the Neuromark_fmri_1.0 template [ 13 ] as the 100 nodes of graph models. The static functional connectivity (FC) was computed as the Pearson correlation coefficients between the time series of each pair of nodes, resulting in a 100 \times 100 correlation matrix for each subject. II-B Proposed Architecture The architecture of SCGT is designed to effectively model the subnetwork structures within the brain by introducing a novel attention mechanism tailored for subclustered graphs. Our model uses static FC correlation features as input to the transformer model. The input graph is constructed with nodes representing brain regions of interest (ROIs) and edge weights representing the correlation coefficients between the ROIs. Positional encodings are incorporated using Laplacian eigenvectors to capture the structural information of the graph. The core component of SCGT is the self-clustering graph attention block, which enables nodes belonging to different clusters of the brain to learn embeddings in a customized manner, addressing the issue of uniform node updates in traditional GNNs and GTs. Unlike standard transformer models that apply the same attention mechanism across all nodes, our approach allows for specialized attention mechanisms for each subnetwork or cluster within the graph. The output node representations from the attention blocks are then aggregated using a concatenation readout function, which combines the features from all nodes to produce a graph-level representation. This representation is then passed through a fully connected layer for the final prediction or classification task. Fig. 1: Overall Architecture of SCGT . The static functional connectivity features are used as input to the self-clustering graph attention blocks. Positional encodings are incorporated using Laplacian eigenvectors. The output node representations are concatenated to form the graph-level representation for prediction tasks. II-C Functional connectivity graph creation The initial phase involves the processing of input from fMRI time series data to create a static FC matrix. Using ROIs

Read the complete text