See DM-12931, which describes the problem and implements a workaround.  A proper solution could be quite a bit more difficult, but is definitely desirable, as we're currently mislabeling masked pixels and probably dropping quite a few pixels we could be keeping.