During the early 1990s when fMRI was becoming more widely used, the neuroimaging community realized that it would be useful to have a common space and coordinate system for reporting results. One of the benefits would be an increase in statistical power for whole-brain analyses; if each of the subjects have BOLD activity in the same general area of the brain, then that signal will be averaged together at a group level, and noise will be canceled out. The other benefit is the ability to place on the same level results by independent researchers. That is, if we all normalize our brain data - i.e., warp and deform them to match a common reference - then the results displayed by one group can be directly compared to those generated by another group.
One of the first templates to be adopted was that of Talairach and Tournoux (1988), a stereotactic atlas based on postmortem dissections of a single subject. In this space, the anterior commissure and posterior commissure - small bundles of fibers sitting below the lateral ventricles that connect both hemispheres - were reoriented so that they would lay straight along a horizontal line. Then, the anterior commissure was designated as the origin, so that any anatomical or functional findings could be reported relative to it; by convention, values went from negative to positive in the following order: Left-to-Right, Back-to-Front, and Bottom-to-Top. For example, a significant fMRI cluster located five millimeters to the left of the anterior commissure, ten millimeters in front of it and six millimeters above it would have MNI coordinates of X=-5, Y=10, Z=6.
Several years later, researchers at the Montreal Neurological Institute (MNI) decided to create a template similar to Talairach space, but which was an average of a normative sample of the population. The first attempt at this was the MNI305 atlas, the number representing the amount of subjects that went into it. The spatial structure was smoother and less defined than the single subject used by Talairach and Tournoux, reflecting the spatial variability of averaging together multiple subjects, and therefore being a more accurate picture of the an effect’s localization. This established the traditional dimensions and resolution of the MNI template, which are 91x109x91 voxels with 2mm^3 resolution.
However, there was a notable dropoff in the coverage of the brain at the bottom of the brainstem and cerebellum, two regions that are often dismissed with a lordly hand by most researchers. The next major version of the MNI template was an average of three sites (MNI, UCLA, and UT Austin) each collecting around 150 subjects and then linearly warping them to MNI space. The result was slightly better definition than the MNI305 brain, and better coverage at the bottom of the brain.
Comparison of different MNI templates. From left to right: MNI305, Colin27, linear MNI152, 40th-generation non-linear MNI152. Note that MNI52 has greater z-axis coverage than MNI305 and that the 40th-generation non-linear MNI152 combines the detail apparent in the single subject Colin27 template with the group representativeness of MNI305 or MNI152. Caption taken from the MNI website: https://mcin.ca/research/neuroimaging-methods/atlases/
Since then, there have been numerous variations and improvements to the MNI152 template, including nonlinear warping which can average together hundreds of subjects into a single template while preserving finer anatomical details. For today’s researcher, most variations of the MNI152 atlas should be suitable for both normalization and for presenting whole-brain results - although there are those who recommend projecting the results onto an average of the individual subject’s anatomical images once they have been normalized, to give a better picture of where the results are localized in your sample.
Today there are also numerous groups creating templates that may be more appropriate for certain populations. For example, there is a frontotemporal dementia template available from McGill University (https://nist.mni.mcgill.ca/mni-ftd-templates/), as well as a pediatric brain composed of 74 subjects from ages 7-14 (Molfese et al., 2014; this atlas comes with AFNI). Using these templates when appropriate can lead to greater statistical power and better localization of your results.