An Augmentation a Day Keeps the MRI Away: Blueprint Augmentation for Brain Graphs

Mariem Touihri, Montassar Ben Dhifallah, Ahmed Nebli

Keywords: Brain graph augmentation, Topology-aware augmentation, Graph Variational Autoencoder

Abstract: Graph augmentations are essential for expanding datasets for classification tasks. Existing graph augmentation techniques often fail to preserve the graph's topological structure. This is particularly important for brain graphs were changes in topology could lead to graph misclassification (i.e., misdiagnosis of brain conditions). This paper introduces Blueprint Graph Augmentation (BluGrAu), a proof-of-concept method designed to address the challenge of augmenting brain graphs derived from MRI data without compromising topological integrity. BluGrAu leverages the novel concept of a "graph blueprint," which is a \emph{transformed} version of an input graph where \emph{essential} topological features of an input graph are preserved. The graph blueprint aims to serve as a template for generating graph variations that retain critical structural properties. By employing a graph neural network-based variational autoencoder (VAE), BluGrAu produces augmented graphs that improve classification accuracy while maintaining a consistent topological structure.

Submission Number: 9