Methods are disclosed for calculating a fat fraction corrected for noise bias of one or more voxels of interest using a magnetic resonance imaging (MRI) system. A plurality of image data sets are obtained each corresponding to NMR k-space data acquired using a pulse sequence with an individual associated echo time tn. A system of linear equations is formed relating image signal values to a desired decomposed calculated data vector having a component such as a water and fat combination having zero mean noise, or having a real fat component and a real water component. A fat fraction is calculated from at least one component of the decomposed calculated data vector. In another embodiment, the system of linear equations is normalized and can directly estimate a fat fraction or a water fraction having reduced noise bias.