Discovering Functional Brain Networks with 3D Residual Autoencoder (ResAE)

Abstract

Functional MRI has attracted increasing attention in cognitive neuroscience and clinical mental health research. Towards understanding how brain give rises to mental phenomena, deep learning has been applied to functional MRI (fMRI) dataset to discover the physiological basis of cognitive process. Considering the unsupervised nature of fMRI due to the complex intrinsic brain activities, an encoder-decoder structure is promising to model hidden structure of latent signal sources. Inspired by the success of deep residual learning, we propose a 68-layer 3D residual autoencoder (3D ResAE) to model deep representations of fMRI in this paper. The proposed model is evaluated on the fMRI data under 3 cognitive tasks in Human Connectome Project (HCP). The experimental results have shown that the temporal representations learned by the encoder matches the task design and the spatial representations can be interpreted to be meaningful functional brain networks (FBNs), which not only include tasks based FBNs, but also intrinsic FBNs. The proposed model also outperforms a 3-layer autoencoder, showing the key factor for the performance improvement is depth. Our work demonstrates the feasibility and success of adopting 2D advanced deep residual networks in computer vision into 3D fMRI volume modeling.