Nonparametric iterative algorithms have been previously proposed to achieve high-resolution, sparse solutions to the bioelectromagnetic inverse problem applicable to multichannel magnetoencephalography and EEG recordings. Using a mmse estimation framework, we propose a new algorithm of this type denoted as source affine image reconstruction (SAFFIRE) aiming to reduce the vulnerability to initialization bias, augment robustness to noise, and decrease sensitivity to the choice of regularization. The proposed approach operates in a normalized lead-field space and employs an initial estimate based on matched filtering to combat the potential biasing effect of previously proposed initialization methods. SAFFIRE minimizes difficulties associated with the selection of the most appropriate regularization parameter by using two separate loading terms: a fixed noise-dependent term that can be directly estimated from the data and arises naturally from the mmse formulation, and an adaptive term (adjusted according to the update of the source estimate) that accounts for uncertainties of the forward model in real-experimental applications. We also show that a noncoherent integration scheme can be used within the SAFFIRE algorithm structure to further enhance the reconstruction accuracy and improve robustness to noise.