A method for maximizing the signal-to-noise ratio (“SNR”) in a combined image produced using a parallel magnetic resonance imaging (“MRI”) technique is provided. The image combination used in such techniques require an accurate estimate of the noise covariance. Typically, the thermal noise covariance matrix is used as this estimate; however, in several applications, including accelerated parallel imaging and functional MRI, the noise covariance across the coil channels differs substantially from the thermal noise covariance. By combining the individual channels with more accurate estimates of the channel noise covariance, SNR in the combined data is significantly increased. This improved combination employs a regularization of noise covariance on a per-voxel basis.