Generative AI enables a new paradigm for brain network construction

Brain networks are a powerful tool for analyzing brain mechanisms and disorders, offering a framework to understand the brain as an integrated system by examining the interactions between different regions. This approach helps map the brain's connectivity, essential for understanding normal function and identifying disruptions in disorders.

However, current brain network construction tools like Pipeline for Analyzing braiN Diffusion imAges (PANDA) and Graph Theoretical Network Analysis (GRETNA) have limitations, including reliance on user expertise, inconsistency in repeated experiments, and time-consuming processes. Generative AI offers significant advantages for brain network generation. It can efficiently learn and generate complex brain networks from limited data, capture unique features for individualized models, and integrate different types of brain imaging data to create more comprehensive and accurate networks.

A research team led by Prof. WANG Shuqiang from the Shenzhen Institute of Advanced Technology (SIAT), Chinese Academy of Sciences (CAS), has introduced a new Diffusion-based Graph Contrastive Learning (DGCL) method for constructing brain networks. This DGCL method serves as a universal tool for identifying key brain connections using generative techniques and could offer valuable insights for neuroscience research.

The study was published in IEEE Transactions on Pattern Analysis and Machine Intelligence on Aug. 15.

By designing a Brain Region-Aware Module (BRAM), the proposed DGCL accurately determines the spatial locations of brain regions through the diffusion process, avoiding subjective parameter selection. Compared to existing methods, the DGCL can learn and capture the unique features of individual brain networks. Subsequently, DGCL employs graph contrastive learning to optimize brain connections by eliminating individual differences in redundant connections unrelated to diseases, thereby enhancing the consistency of brain networks within the same group. Finally, the node-graph contrastive loss and classification loss jointly constrain the learning process to obtain the reconstructed brain network, which is then used to analyze important brain connections.

The proposed DGCL effectively retains and strengthens common connections among similar samples, highlights abnormal brain connections between different categories, and improves the efficiency and stability of brain network construction. Comprehensive results demonstrated that the proposed DGCL outperformed existing tools in terms of both efficiency and prediction accuracy of brain disease.