FLAT-Net: Longitudinal Brain Graph Evolution Prediction from a Few Training Representative Templates

Diagnosing brain dysconnectivity disorders at an early stage amounts to understanding the evolution of such abnormal connectivities over time. Ideally, without resorting collecting more connectomic data time, one would predict disease with high accuracy. At this point, generative learning models from limited can come into play time a single acquisition timepoint. Here, we aim bridge gap between scarcity and prediction by proposing our novel Few-shot LeArning Training Network (FLAT-Net), first framework leveraging few-shot paradigm for connectivity baseline To do so, introduce concept representative connectional templates (CBTs), which encode most centered features (i.e., connectivities) given population networks. Such CBTs capture well heterogeneity diversity, hence they train predictive model in frugal but generalizable manner. More specifically, FLAT-Net starts clustering k clusters using renowned K-means method. Then, each cluster homogenous networks, create CBT, call specific-CBT (cs-CBT). We solely use cs-CBT distinct geometric adversarial network (gGAN) clusters, extract cs-CBTs, gGANs (sub-model) cs-CBT) learn testing stage, compute Euclidean distance subject cs-CBT, select gGAN trained on closest prediction. A series benchmarks against variants excised interpretations showed that proposed FLAT-Net, training strategy, sub-model selection are promising strategies predicting longitudinal alterations only few templates. Our code is available https://github.com/basiralab/FLAT-Net.