This is an MRI question, not a data analysis question per se.

For most MR scanners, the numbers are arbitrary. The signal from the scanner is amplified and adjusted to a convenient level for digitization. The numbers so acquired are then scaled by the reconstruction software to a convenient level -- typically in the 100s to 1000s, so that they can be stored in 16 bit integers (standard image file formats, such as DICOM, don't allow for floating point numbers).

The automatic scanner settings will be somewhat different for different subjects, or even for the same subject on different days. Thus, you can't compare these numbers directly across scanning sessions (the scaling factors normally won't be adjusted during a scanning session).

Inter-regional comparison in the same subject also have a certain level of arbitrariness. This is because the RF coil (essentially the transmit and receive antenna for the radio waves used in MRI) will not have perfectly uniform coverage of the head -- some parts will be illuminated more than others, so the image will be brighter in those parts. For "birdcage" whole-head coils, this illumination factor might be 10-20% (more at high fields like 3-7 Tesla). For this reason, software that tries to segment gray matter from white matter based partly on signal strength must first "uniformize" the image volume to correct for this "shading" effect.

For all the above reasons, and more, it is usual in FMRI to convert signal changes to a percentage of baseline -- baseline in each voxel. In this way, the absolute level of the signal is unimportant.

bob cox