Dear Otto,

For SPM12 "Segment", the TPMs don't have to be in 4D format. However, to make things work (technically at least), you have to create an additional, forth TPM for "non-brain content" so that the TPMs sum up to 1 for any of the voxels, as Christophe has already stated. This can be achieved based on the TPMs for GM, WM, CSF via Imcalc (1 - i1 - i2 - i3). Then specify four tissues, with e.g. 2 / 2 / 2 / 4 number of Gaussians per tissue class, which should avoid the error message "Warning: Matrix is singular, close to singular or badly scaled. Results may be inaccurate. RCOND = NaN." However, the output probably looks rather weird. 

> I've personally tweaked the TPM's in order to include a 7th class into SPM segment without any problem.
While you can indeed add or delete tissues in the GUI SPM12 assumes the TPMs to represent certain tissues / a certain order (e.g. for the clean-up procedures). If you change the order of the TPMs you might no longer get a reasonable segmentation. Thus, in case you want to rely on an age-specific atlas, and with that atlas providing files for three tissue types only, I would suggest to either 1) stick with the "Old Segment". It might be worth to consider some modifications of the internal settings though, as the NIPHD TPMs are sharper than the default SPM TPMs. 

There might be some other differences e.g. due to different skull-stripping quality during template creation, e.g. miclassified meninges are more pronounced in the NIPHD GM TPM along the longitudinal fissure / parietal or the posterior fossa compared to the SPM12 GM TPM, then there's the voxel size, ... Thus, there might well be instances in which an age-inapproriate set of TPMs outperforms age-appropriate TPMs, e.g. because the former is based on a much larger number of subjects (which is usually the case when it comes to pediatric templates), on more advanced methods during template creation, ... this would actually be quite an interesting aspect, but the methods papers usually just conclude that age-approriate TPMs created via xyz are better than their age-inappropriate counterparts, but any of those might be much worse than TPMs created via abc though.

To decrease dependence on TPMs you could thus 2) turn to VBM8/CAT12 segmentation routines, which rely on the TPMs just for some initial preprocessing steps. 

Best

Helmut