(A) A 3D model of a complex of interest. (B) The part of the complex one would like to ignore in masked classification (V1) is shown in yellow. (C) The part of the complex one would like to focus classification on (V2) is shown in red. (D) An experimental particle image is assumed to be a 2D projection of the entire complex in panel A that is affected by the contrast transfer function (CTF), and to which experimental noise (N, shown in grey) has been added. (E) A CTF-affected 2D projection of V1. (F) A CTF-affected 2D projection of V2. Previous approaches to masked classification in RELION (Amunts et al., 2014; Wong et al., 2014) compared experimental particles (panel D), with reference projections of only V2 (panel F). This results in inconsistent comparisons. (G) In the modified masked classification approach, one subtracts the CTF-affected 2D projection of V1 (panel E) from the experimental particle (panel D). This results in a modified experimental particle image that only contains experimental noise and the CTF-affected projection of V2, so that comparison with the reference projection in panel F becomes consistent.