The scalefactors in the headers are essential for the tissue probability maps
to mean anything. I'm not sure exactly what was done, but somewhere along
the way, if the scalefactors are lost, then the TPMs won't work any more.
Best regards,
-John
> If I use custom TPMs in the traditional Analyze format (.hdr+.img pair)
> then SPM5 successfully segments the images, if I first convert the TPMs
> to single-file NiFTI using avwchfiletype from FSL, then the segmentation
> fails totally.
>
> The original TPMs and the NiFTI versions are indistinguishable in SPM5's
> "check reg". SPM5 "display" shows that the originals include an
> intensity scaling factor that seems to have been lost by FSL in conversion.
>
> Ironically, the images I am trying to segment *have* been converted to
> NiFTI (also with FSL), and the SPM default TPMs are NiFTI anyway
> (segmentation with these works). So I would expect this to succeed and
> the Analyze version to fail, if anything.
>
> I attach screen-grabs of spm display for the grey matter TPMs, and check
> reg for the results (showing from left to right: original, native
> grey/white/csf, bias corrected, modulated normalised grey.
>
> My questions: is SPM or FSL (or me!) at fault? How can the (NiFTI)
> segmentation results look so dreadful when the TPMs look fine, and the
> images are themselves NiFTI? And, less directly related, but still
> relevant: why does SPM use such low bit-depth (8bit) TPMs?
|