In dynamic time series imaging, noise varies from image to image, but the underlying structure (although it can vary somewhat) is generally preserved. Techniques that take advantage of this are capable of efficiently denoising such data sets. In this invention, time series images are decomposed into overlapping subblocks of size NxN, with each subblock sampled T times in the time dimension. If each NxNxT subblock is converted to an N-squared by T matrix, promoting low-rank structure in that matrix capture the essential structure while removing noise. The denoised matrices can then be converted back into overlapping subblocks and reassembled into a denoised image series. This technique is easily extendable to 3D data sets sampled over time or to other types of multiple sampling situations. For CT (e.g. perfusion) time series, this technique may allow significant lowering of radiation dose.