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

Abstract

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