Hierarchical Adversarial Connectomic Domain Alignment for Target Brain Graph Prediction and Classification from a Source Graph

Recently, deep learning methods have been widely used for medical data synthesis. However, existing deep learning frameworks are mainly designed to predict Euclidian structured data (i.e., image), which causes them to fail when handling geometric data (e.g., brain graphs). Besides, these do not naturally account for domain fracture between training source and target data distributions and generally ignore any hierarchical structure that might be present in the data, which causes a flat domain alignment. To address these limitations, we unprecedentedly propose a Hierarchical Graph Adversarial Domain Alignment (HADA) framework for predicting a target brain graph from a source brain graph. We first propose to align the source domain to the target domain by learning their successive embeddings using training samples. In this way, we are optimally learning a hierarchical alignment of both domains as we are learning a graph embedding using the previously aligned source-to-target embedding. To predict the target brain graph of a testing subject, we learn a source embedding using training and testing samples. Second, we learn a connectomic manifold for each of the resulting embeddings (i.e., the hierarchical embedding and the source embedding). Next, we select the nearest neighbors to the testing subject in the source manifold in order to average their corresponding target graphs. Finally, using the source and predicted target graphs by our HADA method, we train a random forest classifier to distinguish between disordered and healthy subjects. Our HADA framework outperformed comparison methods in predicting target brain graph and yields the best classification accuracy in comparison with using only the source graphs.