In this work we demonstrate a simple method for
reducing error in k-t under-sampled parallel imaging by subtracting
a dynamic, low-rank time-series estimate prior to un-aliasing reconstruction.
This estimate is generated directly from the under-sampled data by selecting the first $$$r$$$ components of a
singular value decomposition after sliding-window reconstruction,
and removes signal variance that might otherwise contribute to residual
aliasing. This method is motivated by the observation that the highest variance
components in time-series data are typically low-frequency, and well
characterised by a sliding window filter.