Modeling resting state fMRI data via longitudinal supervised stochastic coordinate coding

Abstract

Resting state fMRI (rsfMRI) has been used widely to explore intrinsic brain activities and networks. Although there are a large number of model-driven and data-driven methods that have been employed to model rsfMRI data, it is challenging to model longitudinal rsfMRI data given the time gaps. Currently, sparse dictionary learning (SDL) method has already shown great promise and attracted increasing attention in the rsfMRI research field. The vital advantage of this SDL methodology is that it can identify concurrent brain networks efficiently and systematically. However, the current SDL is not directly applicable to longitudinal rsfMRI data with multiple time points. In response, we propose a longitudinal supervised stochastic coordinate coding (LSSCC) algorithm for longitudinal rsfMRI data analysis. At the first time point, concurrent brain networks are learned and approximated based on the spatial network templates by SDL with l 2 norm. Then, the learned networks at the first time point are transferred to the following time points and the LSSCC is employed to conduct the approximations of functional networks longitudinally. The application of LSSCC on the ADNI-2 longitudinal rsfMRI datasets has shown the effectiveness of our proposed methods.