Decoding Functional Networks for Visual Categories via GNNs

A team of researchers has published a groundbreaking paper, "Decoding Functional Networks for Visual Categories via GNNs," accepted for IEEE ISBI 2026. The work introduces a novel machine learning framework that uses a signed Graph Neural Network (GNN) to model the brain's functional organization when processing visual categories. The model is unique because it accounts for both positive and negative interactions between brain regions, employs a sparse edge mask for interpretability, and uses class-specific saliency to pinpoint critical connections. It was trained on the high-resolution 7T fMRI data from the extensive Natural Scenes Dataset, constructing functional graphs at the parcel level to map brain activity.

The trained AI model successfully decodes distinct functional connectivity states associated with specific visual categories like sports, food, and vehicles. Crucially, it identifies reproducible and biologically meaningful subnetworks that align with known visual processing pathways—specifically the ventral stream (for object recognition) and dorsal stream (for spatial processing). This represents a significant shift from traditional neuroscience methods that focus on activity in individual voxels (3D pixels) to a systems-level, connectivity-based representation. The framework effectively bridges advanced machine learning techniques with core neuroscience questions, offering a new tool to explore the link between cortical organization and human perception.