Abstract
Studying influential nodes (I-nodes) in brain networks is of great significance in the field of brain imaging. Most existing studies consider brain connectivity hubs as I-nodes. However, this approach relies heavily on prior knowledge from graph theory, which may overlook the intrinsic characteristics of the brain network, especially when its architecture is not fully understood. In contrast, self-supervised deep learning can learn meaningful representations directly from the data. This paper proposes a Self-Supervised Graph Reconstruction framework based on Graph-Transformer (SSGR-GT) to identify I-nodes, which has three main characteristics. First, as a self-supervised model, SSGR-GT extracts the importance of brain nodes to the reconstruction. Second, SSGR-GT uses Graph-Transformer, which is well-suited for extracting features from brain graphs, combining both local and global characteristics. Third, multimodal analysis of I-nodes uses graph-based fusion technology, combining functional and structural brain information. The I-nodes we obtained are distributed in critical areas such as the superior frontal lobe, lateral parietal lobe, and lateral occipital lobe, with a total of 56 identified across different experiments. These I-nodes are involved in more brain networks than other regions, have longer fiber connections, and occupy more central positions in structural connectivity. They also exhibit strong connectivity and high node efficiency in both functional and structural networks. Furthermore, there is a significant overlap between the I-nodes and both the structural and functional rich-club. These findings enhance our understanding of the I-nodes within the brain network, and provide new insights for future research in further understanding the brain working mechanisms.
Introduction
Identifying the "hubs" or influential regions of the brain is vital for understanding how we think, learn, and experience neurological disorders. Traditionally, researchers have used graph theory metrics like centrality or rich-club organization to find these hubs. However, these traditional methods rely heavily on prior assumptions and manual feature engineering, which might overlook the brain's intrinsic, data-driven characteristics—especially since our understanding of the brain's "wiring diagram" is still incomplete. Furthermore, supervised deep learning is difficult to apply here because there is no "ground truth" label for what constitutes an influential node.
To address these challenges, this study introduces the Self-Supervised Graph Reconstruction framework based on Graph-Transformer (SSGR-GT). By leveraging self-supervised learning, the framework allows the data to reveal its own structure without human-defined labels, identifying nodes that are most critical for reconstructing the brain's functional state from its structural architecture.
Methodology
The SSGR-GT pipeline consists of three main stages designed to fuse multimodal data and extract node importance through a reconstruction task.
Multimodal Graph Generation
The brain is represented as a graph where nodes correspond to Regions of Interest (ROIs) defined by anatomical atlases. The edges of the graph are defined by structural connectivity, derived from white matter fiber tracking in diffusion MRI (dMRI). The node features are defined by functional connectivity profiles, extracted from signal correlations in functional MRI (fMRI).
The Graph-Transformer (GR-GT) Architecture
The core of the framework is an Encoder-Decoder architecture:
- Encoder: This module combines Graph Convolutional Networks (GCN) to capture local neighbor information with Transformers (Self-Attention) to capture global, long-range dependencies across the whole brain.
- MScore Pool: This novel component assigns a "contribution score" to each node. During training, the model "masks" (hides) low-score nodes, forcing the decoder to reconstruct the network using only the most important information.
- Decoder: Similar to the encoder, the decoder attempts to reconstruct the original functional features of the nodes from the compressed representation.
Reconstruction Task
The model is trained to minimize the reconstruction loss between the original and reconstructed functional feature matrices. Nodes that are most essential for a successful reconstruction—those that consistently receive high scores in the MScore Pool—are identified as I-nodes.
Results
The study utilized the Human Connectome Project (HCP 900) dataset, encompassing 98 healthy adults and various cognitive tasks.
Reconstruction Performance
SSGR-GT demonstrated superior performance in reconstructing brain networks. For resting-state data, it achieved a Structural Similarity Index (SSIM) of 0.82, significantly outperforming baseline models like GraphMAE (0.53). The model also showed high cosine similarity, indicating that the learned representations accurately capture the underlying functional patterns.
Identification and Characteristics of I-nodes
The framework identified a set of 56 "Common I-nodes" that remained influential across almost all cognitive tasks (e.g., emotion, gambling, working memory). These nodes were primarily concentrated in the superior frontal lobe, lateral parietal lobe, and lateral occipital lobe.
Quantitative analysis revealed that I-nodes possess distinct topological attributes:
- High Efficiency: They exhibit significantly higher node efficiency in both functional and structural networks.
- Structural Centrality: They occupy central positions in structural connectivity and have longer fiber connections compared to non-I-nodes.
- Network Vulnerability: "Virtual lesion" experiments showed that removing I-nodes caused significantly more damage to global network efficiency and clustering than removing random nodes or traditional rich-club hubs.
Discussion and Conclusion
The fundamental insight of this work is that reconstruction can serve as a proxy for importance: if a deep learning model requires a specific node to accurately predict the state of the rest of the brain, that node must be a primary coordinator of information. By fusing structure and function, SSGR-GT provides a more holistic view of brain hubs than studies focusing on a single modality.
While the identified I-nodes overlap with traditional "rich-club" hubs, the model also highlighted regions like the sensorimotor cortex and supramarginal gyrus, which are often overlooked by standard graph metrics but appear crucial for task-based integration. Future work will explore the dynamic characteristics of these networks and the potential for I-nodes to serve as biomarkers for neurological disorders.
Summary updated