This work examines the removal of physiologic and measurement noise (i.e. "denoising") of complex-valued EPI timecourse data preceding task-based fMRI analysis. The locally low-rank properties of the EPI data are leveraged with a blockwise singular value thresholding (BSVT) algorithm applied as a preprocessing step. Two EPI datasets (single-band and SMS multi-band) concomitant with task-based finger tapping fMRI exams were preprocessed with BSVT; activation maps were then compared by board-certified neuroradiologists. BSVT denoising of complex-valued fMRI time-course data prior to task analysis improves statistical confidence in activation areas identified by conventional processing or reveals new activation regions under fixed confidence levels.