Automated spectral analysis and estimation of signal amplitudes from magnet
ic resonance data generally constitutes a difficult nonlinear optimization
problem. Obtaining a measure of the degree of confidence that one has in th
e estimated parameters is as important as the estimates themselves. This is
particularly important if clinical diagnoses are to be based on estimated
metabolite levels, as in applications of MR Spectroscopic Imaging for human
studies. In this report, a standard method of obtaining confidence interva
ls for nonlinear estimation is applied to simulated data and short-TE clini
cal proton spectroscopic imaging data sets of human brain. So-called "confi
dence images" are generated to serve as visual indicators of how much trust
should be placed in interpretation of spatial variations seen in images de
rived from fitted metabolite parameter estimates. This method is introduced
in a Bayesian framework to enable comparison with similar techniques using
Cramer-Rao bounds and the residuals of fitted results. (C) 2000 Wiley-Liss
, Inc.