Hi everyone,
In a recent posting
(http://www.mailbase.ac.uk/lists/spm/1999-11/0152.html) Jesper Andersson
mentioned that global intensity normalisation could under certain
cirumstances introduce artefactual activations.
This worried me and I have analysed one of my recent datasets with and
without global scaling normalisation. You can see the results at:
http://www.liv.ac.uk/mariarc/scaletest.html
Basically there were quite marked changes in both positive and
negatively correlated activity. Global intensity normalisation seemed to
reduce the amount of positive activity and enhance the amount of
negative.
So my
questions are:
1) Is this a widely known fact? Personally, I hadnt really considered
that global normalisation could introduce artefacts until Jesper
mentioned it. Once you think it through it's fairly obvious (imagine a
brain in which half the brain is showning postively correlated activity.
After global scaling, the other half will be negatively correlated).
2) Does everybody use global intensity normalisation?
3) The problem is greatest when large areas of the brain are either
positively or negatively activated, such that, presumably, the volume
mean becomes significantly (anti) correlated with the effect of
interest. Does this mean that it is more desirable to have an
experimental designs with a fairly "high-level" baseline condition, so
that the total amount of activity in the brain is minimised, apart from
a few small clusters?
Apologies if I have misunderstood the nature of global scaling
normalisation, or missed a Mailbase posting, but I am relatively new to
the world of SPM.
thanks for your help,
krish
The details of the protocol are on the above web page, but I also
reproduce them here:
------------------------------------------------------------------------
The experiment was a simple AB boxcar design, with A being the task of
interest (Matching to Sample of iconic stimuli) and B being the baseline
condition, which
was a more low-level task in which the subjects simply had to decide
which of six boxes on the screen contained an asterisk symbol.
GE 1.5 T LX/NVi scanner
TR=3 seconds
Total run length=100 volumes (5 minutes)
SOA = 30 seconds
Epoch length=15 seconds
I analysed the data using what I think is a pretty standard first-level
analysis (fixed-effects) of 12 subjects, using an HRF convolved with the
AB profile. After statistical
analysis with a threshold of 0.05 (corrected) I "write filtered" the
results and then visualised them in my own software, superimposed on a
normalised template brain.
I did both the "positive" and "negative" contrasts (1 0) and (-1 0) and
then combined the results in the figures so that orange/yellow
corresponds to clusters
which are positively correlated with task A, and blue/purple/white are
negatively correlated and are, hence, presumably associated with task B.
The top row shows the result when told I SPM99b not to remove global
effects, the bottom row when global effects are removed.
These results show that quite marked change in the balance between
"positive" and "negative" activations can be demonstrated if global
effects are removed.
--
Dr Krish Singh, ([log in to unmask])
Magnetic Resonance and Image Analysis Research Centre, Liverpool
University
Pembroke Place, Liverpool, L69 3BX, UK. Tel 0151 7945645. Fax 0151 794
5635
|