NMR signal contributions from water and fat are separated using a model of the fat resonant frequency spectrum that has multiple resonant peaks. The relative frequencies of the multiple fat spectrum peaks are known a priori and their relative amplitudes are determined using a self-calibration process. With the determined relative amplitudes of the fat spectrum peaks, acquired NMR signals are modeled. Using this model and NMR signal data acquired at a plurality of echo times (TE), the signal contribution from multiple fat spectrum peaks is separated from the acquired NMR signal data. A combined image is alternatively produced from weighted contributions of the separated water and fat images. Additionally, a more accurate estimation of the apparent relaxation time and rate (T2* and R2*, respectively) is alternatively performed.