Brain volume prediction based on rs-fMRI time series

Rs-fMRI time series and volumes of different brain regions possess rich information about the functionality and the structure of the brain, respectively. Recently, artificial neural network emerges as a powerful tool to perform machine learning tasks, which empowers the exploration of the direct relationship between the two. We introduced the deep learning based method and proposed a 1-D convolution operation based deep structure that utilized the spatial and temporal features of the rs-fMRI time series to predict the volumes of different brain regions. The proposed method outperformed the baseline fully-connected neural network, in terms of mean square error. This indicated that the proposed method can utilize temporal and spatial information from rs-fMRI time series better than the fully- connected counterpart.