3D-EPI is beneficial for real-time fMRI applications because the acquisition can be accelerated in both the in-plane and through-plane phase-encoded directions significantly increasing temporal resolution. Fully-sampled reference data are required in order to map coil sensitivities and unfold the aliased images that result. Acquiring these data with a segmented EPI readout matches the EPI-related distortions and dropouts between the un-accelerated reference data and the accelerated fMRI data. However, systematic differences in eddy-current-related phase between the segments can introduce errors during frequency offset correction (employed to correct for slow frequency drifts) unless accounted for. Here we present a simple correction scheme that greatly improves the tSNR of the time series.