I was asked by Andriy Fedorov to report this here. In the case where data with a small number of slices is passed to N4ITKBiasFieldCorrection, the default shrinkfactor of 4 causes the algorithm to converge nearly immediately with no change in the image (and a bias field of all ones). Reducing the shrinkfactor allows it to proceed as normal. It's not clear whether the calculated field is actually worth anything with a partial brain, but as suggested by Andriy, it may be worth a warning when the chosen shrinkfactor is too large for the image size. (It also might be nice to allow it to subsample differently for each dimension)