Joint Estimation of Neural Events and Hemodynamic Response Functions from Task fMRI via Convolutional Neural Networks

Abstract

Joint decomposition of functional magnetic resonance imaging (fMRI) time series into time courses of neural activity events and hemodynamic response functions (HRF) can enable new insights into functional connectivity, task activation, and neurovascular coupling in health and disease. Current methods for this problem handle time series of either temporally isolated events or extended blocks of continuous events but not both; and they constrain the HRF to one specific functional form. We propose to use an autoencoder and a convolutional neural network (CNN) to overcome these challenges. The autoencoder uses convolutional neural networks to reconstruct the fMRI time series while estimating the neural event time series. The CNN estimates the HRF as the convolutional filter that, when applied to a binarized version of the neural event time series, best reconstructs the fMRI time series. When applied to synthetic data and data simulated by the STANCE fMRI simulator, the method estimates ground-truth neural events and HRFs more robustly than competing methods. When applied to real-world fMRI data, the method identifies temporally isolated, continuous, or mixed neural events that correspond to experimental conditions more closely than competing methods. The flexibility and computational power of machine learning techniques enable the accurate capture of diverse HRFs and neural event time series from fMRI data.