Adversarial Connectome Embedding for Mild Cognitive Impairment Identification Using Cortical Morphological Networks

Abstract

Cortical Morphological Networks provided unprecedented insights into connectional brain alternations in patients diagnosed with mild cognitive impairment (MCI) and, in combination with deep learning techniques, they can further be utilized to build computer-aided MCI diagnosis models. In this paper, we introduce Adversarial Connectome Embedding (ACE) architecture, which is rooted in graph convolution and adversarial regularization to learn relevant connectional features for MCI classification. Existing connectome-based embedding methods for examining the healthy and disorder brain connectivity generally rely on vectorizing the connectivity matrix and use typical Euclidean embedding methods (e.g., principal component analysis), which work best in Euclidean spaces such as images. On the other hand, a connectome, which is modeled as a brain graph or network, lies in a non-Euclidean space. Hence, the connectome vectorization might cause losing its topological structure which can be leveraged to boost brain graph classification for diagnosis. To fill this gap, we leverage geometric deep learning, a nascent field which extends deep Euclidean feature representation learning to non-Euclidean spaces. First, we propose to use a geometric autoencoder with graph convolutional layers to learn a latent brain connectivity representation (i.e., embedding) that exploits the connectome topology. Secondly, we utilize an adversarial regularizing network which forces the learned latent distribution to match the prior distribution of the connectomes. Finally, we feed the adversarially regularized latent connectome embeddings to train a linear classifier for diagnosing MCI patients. ACE achieved the best classification results across different connectomic datasets for MCI versus Alzheimer’s disease classification in comparison with typical graph embedding techniques.