Spherical deconvolution is an elegant method by which the orientation of crossing fibers in the brain can be estimated from a diffusion-weighted MRI measurement. However, higher resolution of fiber directions comes at the cost of higher susceptibility to noise. In this study, we describe the use of linear regularization of the fiber orientation distribution function by Damped Singular Value Decomposition. Furthermore, the degree of regularization is optimized on a voxel-by-voxel basis with no user interaction using Generalized Cross Validation. We find, by simulations, that regularization can improve the reliability of fiber orientation determination when the signal-to-noise ratio is low. Simulations and in vivo measurements indicate that spurious peaks of the fiber orientation distribution function in regions with low anisotropy largely disappear when regularization is introduced. The methods examined are fast enough to be used on a routine basis with diffusion MRI data sets and may improve estimation of water diffusion properties in heterogeneous white matter and boost reliability of fiber tracking through regions of brain with complex fiber geometry.