BGCSL: An unsupervised framework reveals the underlying structure of large-scale whole-brain functional connectivity...

Inferring large-scale brain networks from functional magnetic resonance imaging (fMRI) provides more detailed and richer connectivity information, which is critical for gaining insight into brain structure and function and for predicting clinical phenotypes. However, as the number of network nodes increases, most existing methods suffer from the following limitations: (1) Traditional shallow models often struggle to estimate large-scale brain networks. (2) Existing deep graph structure learning models rely on downstream tasks and labels. (3) They rely on sparse postprocessing operations. To overcome these limitations, this paper proposes a novel framework for revealing large-scale functional brain connectivity networks through graph contrastive structure learning, called BGCSL.

Unlike traditional supervised graph structure learning methods, this framework does not rely on labeled information. It consists of two important modules: sparse graph structure learner and graph contrastive learning (GCL). It employs dynamic augmentation in GCL to train a sparse graph structure learner, enabling it to capture the intrinsic structure of the data.

We conducted extensive experiments on 12 synthetic datasets and 2 public functional magnetic resonance imaging datasets, demonstrating the effectiveness of our proposed framework. In the synthetic datasets, particularly in cases where node features are insufficient, BGCSL still maintains state-of-the-art performance. More importantly, on the ABIDE-I and HCP-rest datasets, BGCSL improved the downstream task performance of GCN-based models, including the original GCN, dGCN, and ContrastPool, to varying degrees.

Our proposed method holds significant potential as a valuable reference for future large-scale brain network estimation and representation and is conducive to supporting the exploration of more fine-grained biomarkers.