Which method should I use?

There is no gold standard for MR intensity normalization — the right choice depends on your data and downstream task. This guide is the practical summary; see How the methods work for the math.

Decision guide

One T1-w image per subject (the common case) : Start with FCM (inorm.fcm). It scales the white-matter mean to 1 using fuzzy tissue memberships and is robust in practice. If results look odd in non-WM tissue (check with a histogram plot!), try LSQ across your set, which balances CSF/GM/WM means jointly.

No T1-w image, or non-brain anatomy : Use zscore (any anatomy, any modality) or kde with the right --modality so the correct histogram peak is used.

A dataset for machine learning (train/test split) : Use nyul: fit the standard histogram on the training set, save the transform, and apply the same transform to validation/test/inference data. Histogram matching tends to equalize the whole distribution, not just one landmark — often what ML pipelines want. RAVEL and LSQ transforms are likewise savable.

Multi-scanner/multi-site study, co-registered images, visible batch effects : Use ravel: WhiteStripe + removal of latent technical factors estimated from CSF control voxels. It's the most involved (needs co-registration, batch-only) and the most aggressive at removing scanner effects.

Principled per-image standardization with a paper behind it : Use whitestripe: z-score within the normal-appearing white matter.

Modality matters

kde and whitestripe anchor on a tissue peak of the smoothed foreground histogram, and which peak is the right one depends on the modality:

modality	peak used	rationale
t1	last	WM is the brightest of the three tissues
t2, flair	largest	global maximum of the histogram
pd, md	first	lowest-intensity tissue peak
other	last (override with --peak)	escape hatch for non-standard data

For contrast-enhanced or otherwise non-standard images, read How the methods work, pick the peak that matches your tissue of interest, and pass --peak / peak= explicitly. Then validate with a histogram plot before batch-processing hundreds of scans.

What every method assumes

• Brain masks (or skull-stripped input) for everything except zscore and nyul, which tolerate full-head images better. Bad masks are the top cause of bad results. The package fails loudly with an actionable message, but garbage in is still garbage in.
• All brain-specific methods (fcm, kde, whitestripe, lsq, ravel) assume three tissue classes (CSF/GM/WM) exist. Heavy pathology, infants, or non-human primates can violate that.
• Population methods assume all input images share one modality and comparable acquisition.