Dear Ananth,


I'd say segmentation for 1) VBM and 2) for normalisation purpose are two different issues. Anatomical volumes might be sufficient for preprocessing aspects but "too bad" for VBM at the same time, although I don't know of any "rules" or guidelines. Reviewers might have a different opinion of course, but here some subjective thoughts:

1) If you want to conduct VBM then any artefact (like the ringing) is going to bias / lower the quality of VBM to some extent, at least locally. Often it is also problematic because artefacts might be more frequent in one group (e.g. head motion in patients), so differences in grey matter might actually be differences in data quality. In a strict way you would have to exclude any of these volumes, as there are visible artefacts and thus they probably don't meet your expectations. However, this is rather difficult when it comes to volumes with e.g. lower SNR overall. The VBM8 toolbox has an option to detect outlier volumes based on sample homogenity, thus providing some objective criterion. It's a little tricky of course and you should still check the data yourself, as volumes with clear and possibly severe local artefacts might still be within the "normal" range and other volumes with good data quality but maybe some anatomical anomalies (e.g. large sinuses in healty or maybe large atrophy in patients) might fall outside. To avoid any confusion due to differences in head structures like the neck you might test the sample homogenity of the skull-stripped versions (or c1* for VBM).

2) For segmentation-based normalisation the data quality should be less important, as you're just interested in the normalisation part. As far as I remember some smoothing is applied during coregistration, so artefacts like the ringing should actually be of a minor issue here. The coreg is not perfect anyway due to its nature (between-modality registration of two images with quite different resolutions, different geometric distortions, different dropouts) and manual adjustments might be necessary, especially if you don't cover the entire cortex during the EPI session. In that case the structure might be shifted a few mm too much along z after coreg. I once played with some animal fMRI data with extreme dropouts in the EPIs, the automatic coregistration was worthless in that case. So just try, also with the motion-corrupted volumes, and then check the outcome.

Concerning the segmentation itself, I'm not sure about the data in your case, but with motion-corrupted volumes the border between grey and white matter is typically very blurry. Voxels are going to be misclassified based on the intensity then, but the normalisation part / deformations tends to be okay in my experience (although this is a subjective impression of course). But make sure this holds for your data as well.

In any case you should definitely use the same preprocessing pipeline within a study. The different normalisation options (segmentation-based, T1 template, EPI template) lead to somewhat different outcomes, also concerning the global shape, even if we regard all of them as MNI space afterwards. In my opinion EPI template normalisation is equivocal to some extent anyway, as the distortions and signal dropouts might be quite different in your raw data compared to the template, and the distortions might also vary between subjects depending on slice positioning. For example if you have large dropouts in ventromedial frontal cortex, then the more dorsal aspects might (incorrectly) be warped downward. Same holds true if you don't cover the entire cortex due to limitations from slice thickness * number of slices. Assume the topmost part of the cortex was outside the filed of view. With EPI normalisation the covered parts are going to be warped out-/upwards to some extent, and your normalised data might look as if you've covered the whole brain.

Thus, if the segmentation or T1 template normalisation does not work well for your bad volumes, either turn to mean to EPI template normalisation for the whole study  or discard the functional series of affected subjects (or try to rescan these subjects). Typically the data loss shouldn't be that dramatic as subjects with lots of head motion during the anatomical scan probably move around during the (much more disturbing) EPI session a lot as well.


Hope this helps a little,

Helmut