Data-Efficient Brain connectivity Analysis via Dual Meta-learning for Brain Disorder Detection

By modelling complicated brain area interactions, brain networks can shed light on human cognition/behavior mechanisms and uncover biomarkers to brain illnesses. Deep learning methods like graph neural networks have become popular recently for brain connectivity analysis, but brains have limited amounts of connectivity, which forces models to overfit. We built DuMeta-connect, a data efficient brain connection analysis technique based on meta learning, which learns by distilling the experience of many learning tasks to enhance future learning performance. Unlike standard brain connectivity analysis, we generate many brain connectivities for the same patients and treat each classification as a source task. We apply a dual meta learning approach that learns a brain connectivity agnostic feature extractor to encode generic brain representation and a feature fusion well initialized by source tasks to learn fused representation that is aligned to task. The well initialized feature fusion is meta tested for target tasks. We evaluated the proposed resting-state functional magnetic resonance imaging approach for detecting ASD in locations from the publically accessible ABIDE dataset. Experiments suggest that our strategy performs better and more consistently than many state of the art approaches.