<snippets> from eelke code:
USRCXXFLAGS = -std=c++11
mist/mist: ${OBJS} mist/mist.o
${CXX11} ${CXXFLAGS} ${LDFLAGS} -o $@ $^ ${LDLIBS}
also to enable a c++11 ready bash session on centos 6
scl enable devtoolset-2 bash
> On 24 Nov 2016, at 10:49, Jesper Andersson <[log in to unmask]> wrote:
>
> Dear Gabor,
>
> I am curious how big differences you see?
>
> It is true that eddy is not deterministic (because of the random selection of nvoxhp voxels to base the GP hyperparameter estimation on), but my experience is that those differences are very small. In the outlier paper we ran lots of tests with simulated data which had unique noise realisations, and even in that case (where the difference was both from completely different noise and from the random selection) the realisation-to-realisation FA differences were very small.
>
> My gut feeling would be to say “don’t worry about it” and that if you still worry to increase nvoxhp to for example 5000. But that of courses hinges on your run-to-run differences being of the same order of magnitude as those we see.
>
> I wouldn’t run multiple eddys.
>
> Jesper
>
>
> On 23 Nov 2016, at 17:30, Gabor Perlaki <[log in to unmask]> wrote:
>
>> Dear Jesper,
>>
>> I think our differences is more related to the parameter --nvoxhp and the random component in eddy.
>>
>> 1, Will a bigger --nvoxhp result in more correct my_eddy_output.nii.gz and FA/MD values?
>>
>> 2, Since eddy_cuda is quite fast on our machine, we also have the opportunity to run eddy several times (e.g. 10) with the default parameters (--nvoxhp=1000), run dtifit 10 times and average the 10 FA maps.
>>
>> Which option is better to increase the reliability of our final FA/MD maps?
>>
>>
>>
>>
|