The registration method described above in sections 4 and 5 has been implemented in C++ and is called FLIRT -- FMRIB's Linear Image Registration Tool. This program has undergone extensive trials over several months, being used by various researchers including trained neurologists, psychologists and physiologists. During this time it has been used to perform many thousands of registrations in the context of FMRI analysis and structural studies [Smith et al., 2000].
Feedback from the users has been positive, with the vast majority of registrations producing acceptable results and only a few cases of failure, or visually unacceptable registrations. Of the failures reported it was found that the other registration methods available (see below) also failed to find an acceptable registration. Furthermore, it was found that these volumes were usually difficult to manually register as they often had unusual features such as particularly enlarged ventricles. Consequently, these cases are ones where the main problem appears to be the transformation space. That is, a good affine registration does not exist. It would require the use of higher order transformations (non-linear warpings) to achieve a good solution.
One aspect that was highlighted during these trials was that, as reported before [Woods et al., 1993], in cases where the scalp or skull appeared in one image but not the other, better results were sometimes obtained by ``stripping'' the scalp and skull from the image. Without doing this stripping a typical registration error that occurred was that the outer brain surface would be aligned with the scalp/skull surface in the other image. That is, the scale was incorrectly set. This is a direct result of the cost function having similar values for this situation and the ``correct'' scale, since in one case the gap between the scalp/skull and brain is matched poorly, whilst in the other case the scalp/skull itself is matched poorly. Therefore, it should be kept in mind that such skull/scalp stripping should still be performed, either manually or automatically, in order to get the best results.