Figure 2
From: Attention-based fusion of multiple graphheat networks for structural to functional brain mapping
Proposed A-GHN architecture for learning SC-FC mapping using multi-scale GraphHeat networks (GHN) along with attention mechanism. A Laplacian matrix is computed from the structural connectivity matrix (SC) input in step 1. Multiple heat kernel matrices are obtained using m different diffusion scales and fed to the individual (A-GHN sub-model) in step 2. In step 3, an attention module is introduced to learn the attention scores corresponding to A-GHN sub-models. A Softmax linear combination of the outputs \(\Psi _{\gamma _i}\) yields the predicted functional connectivity (\(C_f\)), which is compared with the ground truth empirical FC in step 4.
