Since the spherical-deconvolution (SD)-based intravoxel fiber-orientation distribution (FOD) estimation method is highly sensitive to noise, a non-convex regularized SD method is proposed. It constructs a non-convex spatial regularization based on the FOD similarity between neighboring voxels and resolves the non-convex regularized SD problem using the modified Richardson-Lucy algorithm. The simulated results based on data in two tensors model and HARDI (high-angular-resolution diffusion imaging) model show that, compared with the conventional SD and total-variation regularized SD methods, the proposed method generates FODs with a lower mean angular error (reduced by 52% and 9%, respectively) and exhibits better noise immunity and detail preservation of the fiber orientations.