Schematic of the proposed embedding framework. (a) Identification of GyralNet and 3HGs. a1 White matter cortical surface mapped by gyral altitude; a2 segmentation of gyral crest (white regions) from sulci basins (color regions); a3 connecting the gyral crest regions into a completed graph by tree marching (black curves). A magnification view of the circled patch is displayed between a2 and a3; a4 pruning the redundant branches to preserve the main trunk of the graph (black curves)—GyralNet; a5 identification of 3HGs (labeled by green bubbles). (b) 3HG’s multi-hop feature encoding. b1 Parcellating the entire cortex into 75 ROIs via Destrieux Atlas and assigning each 3HG with an ROI label as node feature; b2 Numerically representing each ROI label by one-hot encoding; b3 by considering multi-hop neighbors, 3HGs are encoded by multi-hop features. (c) The proposed learning-based embedding framework (details can be found in Section 2.4).