Magnetic Resonance Imaging (MRI) uses the resonance of the nucleus of
chemical species to generate images of their spatial distribution. In
medical MRI, a simple model considers tissue as made up of only water
and fat. Most of the clinically relevant information is in the water
signal and the fat signal is considered clutter to be suppressed. The
separation of water and fat based on the difference of their resonance
frequencies using multiple images provides a robust method for fat
suppression in areas where the magnetic field is inhomogeneous. In this
talk, we will show how to propagate the uncertainty due to imperfections
of the magnetic field into our estimate of water and fat. The
optimization of data acquisition based on the Cramer-Rao bound (CRB) for
this nonlinear problem leads to new optimal solutions which do not arise
when the magnetic field is assumed to be homogeneous. A reconstruction
based on maximum likelihood estimation allows us to achieve the CRB for
realistic noise levels which is verified by Monte Carlo simulations,
experimentally and clinically. Our acquisition and reconstruction is
part of a method for chemical species separation currently used by
General Electric MRI scanners.