Abstract
Multi-echo independent component analysis (ME-ICA) has been demonstrated to improve sensitivity and reliability of task functional magnetic resonance imaging (fMRI) data and, in particular, motor-task data with inherent task-correlated head motion. However, previous work has shown that an overly aggressive ME-ICA denoising approach may unintentionally remove task-related signal, while a more conservative approach may not effectively mitigate noise. While the effects of varied implementations of ME-ICA on signal and noise characteristics have been tested thoroughly in breath-hold data, the effects of similar modeling decisions have not been studied in motor-task data, which present with a more localized neural response. Here, we tested and compared the impacts of three analysis methods using rejected ME-ICA components as regressors in subject-level modeling: Aggressive (simple inclusion of ME-ICA regressors), Moderate (excluding task-correlated ME-ICA regressors from the model), and Conservative (orthogonalization of ME-ICA regressors to the base model and accepted ME-ICA components). We applied these methods to data from healthy and multiple sclerosis populations that included performance of hand-grasp, shoulder-abduction, and ankle-flexion tasks. We found that when the amount of head motion and its correlation with the task was high and the expected task-evoked signal was relatively low, the Conservative method led to significantly higher activation, t-statistics, and test-retest reliability in motor regions compared to the Aggressive and Moderate methods. Future motor-task studies may wish to implement similar models to prevent loss of motor signal, while still mitigating the effects of task-correlated head motion.