BrainNet-GAN: Generative Adversarial Graph Convolutional Network for Functional Brain Network Synthesis from Routine Clinical Brain Structural T1-Weighted Sequence.

Functional brain network (FBN) derived from functional Magnetic Resonance Imaging (fMRI) has promising prospects in clinical research, but fMRI is not a routine acquisition data, which limits its popularity in clinical applications. Therefore, it is imperative to generate FBN based on routine clinical structural MRI brain network. In this study, a BrainNet-GAN model was proposed for generating FBN from radiomics-based morphological brain network (radMBN) derived from routinely acquired T1-weighted image (T1WI). BrainNet-GAN integrated two Multi-Channel Multi-Scale Adaptive (MultiAda) generators and two (Local_to_Global) discriminators. In the generator, Graph Convolutional Network (GCN) was used inside each channel to aggregate multi-scale information between direct or indirect neighbors of nodes, and the output of each channel was adaptively fused through several sets of learnable coefficients; In the discriminator, Multi-channel GCN was used to aggregate local nodes information, and a feature selection module was designed to establish correlations between feature maps at different channels. Additionally, a Multi-Angle Multi-Constraint (MAMC) loss function was proposed, which could guide the learning process of the model from different aspects. Experiments with 2116 subjects in two publicly available datasets showed that BrainNet-GAN model exhibited promising performance on the task of generating FBN. Meanwhile, the individual-level brain network visualization was displayed with high consistency in generated FBN and target FBN. Further, the Top 10 brain regions identified by four graph-theory analysis metrics also exhibited with consistency. The proposed BrainNet-GAN model demonstrated superior performance in generating FBN based on radMBN, which could facilitate the application of FBN in clinical practice.