In single-particle reconstruction, a 3D structure is reconstructed from a large number of randomly oriented 2D projections, using techniques related to computed tomography. Unlike in computed tomography, however, the orientations of the projections must be estimated at the same time as the 3D structure, and hence the reconstruction process can be error-prone, converging to an incorrect local optimum rather than the true 3D structure. In this paper, we discuss and further develop a maximum-likelihood approach to reconstruction, and demonstrate that this approach can help avoid incorrect local optima for both 2D and 3D reconstructions.
Marshall W. Bern, Jindong Chen, Hao Chi Wong