Hi,
On 6 Jun 2007, at 20:33, Antonios - Constantine wrote:
> Dear Steve -fsl users
>
> Thank you very much for the information about randomise..Still i
> need some
> more help in order to explain the results that i have after running
> randomise with 8 controls and 14 Parkinsons subjects. I used a c
> threshold
> of 1 and the max amount of permutations (more than 300.000
300000! how long did that take! normally 5000 should be enough :)
the program tells you what the _maximum_ number of iterations is - in
general if this is greater than 5000 you don't need more than 5000.
> ) as the randomise
> tool suggested for the best accuracy in my case.More specific :
>
> randomise -i flow_all_subjets.nii.gz -o sienar -m
> /usr/share/fsl/etc/standard/avg152T1_edges -d design.mat -t
> design.con -n
> 319770 -V -c 1
>
> Contrast1: Parkinsons>Controls
> Contrast2:Controls> Parkinsons
> Contrast3:Parkinsons mean
> Contrast4:Controls mean
>
> I uploaded the results of the randomise test and i would really
> appreciate
> if you could find some time and have a look at them. Ref.number:
> 942183
Nice - your data looks fairly convincing I think.
The largest cluster is nearly significant - it's hard for clusters to
reach significance when the cluster-forming threshold is so low (1 in
your case).
I re-ran at -c 2.5 (a more "normal" value) and the clusters around
the ventricles and the middle temporal gyrus are now significant
(well, 0.94).
I'm afraid there isn't a good way to know what a good cluster-forming
threshold is.....that's a downside of cluster stats.
You might also want to play with the extent of spatial smoothing in
the siena_flow2std script which will affect these results.
> Firstly i checked the max_tstats 1&2 (in fslview dynamic range
> 0.949:1) and
> there weren't any significant voxels there. Then i checked the
> maxc_tstats
> 1&2 (in range 0.949:1) for clusters but i also didn't see any
> significant
> clusters. That means that there is no difference between the
> controls and
> the patients?
see above
> As you advised me i checked after that the tstat1 and tstat2 which
> were
> filtered also in range [0.5:3] as it was in your example even though i
> didn't understand exactly why i should use this filtering range and
> not for
> instance [0.1:5].What's the meaning of filtering our tstats 1&2
> data in a
> specific range? What about the intensities of these tstats?
By looking at the raw tstat images you are seeing the unthresholded
statistical strength of the effect. If you just load tstat1, make its
colourmap Red-Yellow, and click on the (i) and turn on the button
underneath the colourmap selector so that you can turn on the
secondary colourmap to be used for negative values, and set that to
Blue; then if you set the display range you get to see negative
values less than -1 in blue and greater than 1 in Red-Yellow. If you
increase the minimum intensity display range to say 2 instead of 1
then you can see what the clusters look like for -c 2, _before_ they
are turned into p-values.
> The bigger their
> magnitude the more the evidence of atrophy/growth between the
> Patients and
> the Controls in this area??By reducing their range with fslview
> what do we
> succeed?
>
> There are evidence in both tstat1 and tstat2 about atrophy (the
> value of
> which, -positive or negative- will be evaluated with tstats3&4).
> So, correct
> me if i'm wrong, we need both tstat1 and tstat2 in order to see all
> the
> areas where there's atrophy between Patients/Controls.If we check only
> tstat1 or tstat2 then we miss a lot of information about atrophy/
> growth
> areas, right?
Well, you can use just tstat1 to see both directions, as it includes
negative values, but to test for statistical signicifance, with the
maxc cluster pvalues, you need both maxc1 and 2, as maxc1 only tests
controls<patients and maxc2 tests controls>patients. You have much
more signal in contrast as you'd expect, I guess your patients have
more atrophy than the controls.
You can actually get a very nice clear picture of the whole story by
loading up the full 4D data into FSLView as the first thing you load.
So run:
fslview flow_all_subjets sienar_tstat1
then go to an interesting voxel (say near the ventricles) where
tstat1 is strongly negative, and then turn on the timeseries view.
You can see that consistently the patients have atrophy (negative
values around -0.1) and controls are much closer to 0. Looking at it
this way you don't really need tstats3 and 4 to help interpret the
differences seen in the first two contrasts. (though 3 and 4 are
simple ways of summarising the within-group means).
> How these statistics (tstat1&2)show evidence of atrophy/growth, while
> maxc_tstats and max_tstats don't give us that information? It's
> because
> tstats shows only evidence of atrophy while maxc_tstats and
> max_tstats shows
> the significance of these evidence?
>
> And the final step is to check tstat3 and tstat4 in order to find out
> whether each atrophy that was detected with tstat1 and tstat2 is
> positive or
> negative.. In this step i didn't know what range i should use in
> order to
> filter with fslview the data, so i used randomly the range [0.5:3]
> (the
> range that was in tstats1&2).But then i didn't know which were the
> evidence
> in tstat3&4 that shows positive/negative values for the
> atrophy...Is it the
> intensities?positive intensities in tstat3 shows atrophy in Parkinsons
> patients and negative once shows growth, while positive intensities in
> tstat4 shows atrophy in Controls and negative once shows growth on
> them?
> What about the magnitude of the intensities in tstats 3&4? the
> bigger the
> absolute value of the intensity in a voxel the bigger the state of
> atrophy/growth in this voxel?
>
> And finally since tstats are raw data, before randomisation, that
> means that
> we don't have any idea about the significance of the evidence that
> we see
> with these data..right?So why do we use them to extract information
> about
> atrophy? It's only some basic evidence,but we don't know if it's
> important
> evidence, only p-values will show us if these evidence are of
> importance...
You have the tstats partly because they are an intermediate stage in
the analysis that ends up making p-values, and partly because looking
at the unthresholded tstats is a good way of seeing what's in your data.
Hope this is making sense?
Cheers, Steve.
>
> I'm sorry for asking again so many questions but i don't really
> know how to
> explain the results that i had after running randomise and i'm not
> sure if
> these results that i already uploaded for your consideration shows
> some
> statistical significant evidence about atrophy or growth in my
> groups of
> study (Parkinsonians -Controls), so i would really need your
> opinion on that.
>
> thanks a lot again for your time and effort.
> Antonios-Constantine Thanellas
------------------------------------------------------------------------
---
Stephen M. Smith, Professor of Biomedical Engineering
Associate Director, Oxford University FMRIB Centre
FMRIB, JR Hospital, Headington, Oxford OX3 9DU, UK
+44 (0) 1865 222726 (fax 222717)
[log in to unmask] http://www.fmrib.ox.ac.uk/~steve
------------------------------------------------------------------------
---
|