[Mrtrix-discussion] Number of tracks
Kerstin Pannek
uqkpann1 at uq.edu.au
Thu May 12 16:44:50 PDT 2011
Hi Luis
> According to the help in mrtrix for 30 directions wouldn't be 6 the
> ideal value for lmax?
You are right, the documentation says lmax=6 for 30 directions. I had convergence errors with lmax=6 for my data (neonates, not adults), so I reduced it to lmax=4, and by now I automatically think 30 directions -> lmax 4; sorry
> How do you decide when to stop adding streamlines, you do a visual
> inspection or use some voxel-wise mathematic?
That depends on your application. If you want to use the extracted tract as volume of interest to extract, for example, FA, visual inspection might be good enough. If you want streamline number, streamline length, etc, you would calculate the stdev across those numbers. If you final application is about maps, then you do voxel-wise stdevs.
> I expect some variability among the groups I'm going to compare. So I
> suppose the best idea is to check this variability in each of the
> groups. Would it be better to have one number of streamlines for each
> group, or picking the maximum of all of them and have just one value
> number for all of them? Because ideally the only problem would be a
> longer computational time. But once you have enough streamlines to
> avoid variability adding more is not a problem, is that correct?
There shouldn't be too much variability in the number of streamlines you need between the groups. There will, however, be variability in your _final_metric_ across participants - and hopefully significant differences between groups.
> I also have two sets of data, images were acquired in different
> places. One of the sets is 1.25x1.25x2.5 the other is 2.5x2.5x2.5. My
> doubt if it's possible any comparison between this sets.
I doubt it is possible to compare those sets... If you want to try and compare them, and want to use the seeds per voxel approach (i.e. different number of streamlines for each participant), keep the number of seeds per cubic mm constant. The other option is to keep the overall total number of streamlines constant for all participants.
> What's the nifti format of the eigen vector
> file produced by mrtrix? It has three dimensional data so I suppose
> one volume for each component, but what's the order. X.Y.Z component
> as obvious?
I believe it is x.y.z
Cheers
Kerstin
On 12/05/2011, at 6:40 PM, Luis Morís wrote:
> Hi again,
>
> Thanks for your answer Kerstin.
>
> I'll go the way you've described me.
>
> I have some more questions.
>
> According to the help in mrtrix for 30 directions wouldn't be 6 the
> ideal value for lmax?
>
> How do you decide when to stop adding streamlines, you do a visual
> inspection or use some voxel-wise mathematic?
>
> I expect some variability among the groups I'm going to compare. So I
> suppose the best idea is to check this variability in each of the
> groups. Would it be better to have one number of streamlines for each
> group, or picking the maximum of all of them and have just one value
> number for all of them? Because ideally the only problem would be a
> longer computational time. But once you have enough streamlines to
> avoid variability adding more is not a problem, is that correct?
>
> I also have two sets of data, images were acquired in different
> places. One of the sets is 1.25x1.25x2.5 the other is 2.5x2.5x2.5. My
> doubt if it's possible any comparison between this sets.
>
> My ideas are:
>
> a) Corregister. I'm considering moving both patients to 2x2x2, my
> problem is that I don't think making any kind of displacement in DWI
> images is a good idea, because it'll probably give some problems as
> they are related to the vector directions.
>
> b) As The first set has 4 times the number of voxels of the second
> set, either place 4 times more seeds on the second set or place 4
> times less seeds in the first set. But It will probably introduce some
> bias.
>
> c) Find a common maximum number of streamlines and compare them
> without any further consideration. Simply check variability between
> sets.
>
>
> And one last question. What's the nifti format of the eigen vector
> file produced by mrtrix? It has three dimensional data so I suppose
> one volume for each component, but what's the order. X.Y.Z component
> as obvious?
>
> Thanks in advance for your help,
>
> Luis.
>
>
> On Thu, May 12, 2011 at 12:49 AM, Kerstin Pannek <uqkpann1 at uq.edu.au> wrote:
>> Hi Luis
>>
>> Are you doing tensor tractography? You should be able to use constrained spherical deconvolution with your data. I have some 30 direction data with b=1000, and use csdeconv with -lmax 4.
>>
>> Medinria uses deterministic tractography (correct me if I'm wrong). With deterministic tractography, there is only one possible streamline for every seed point. With probabilistic tractography, there is a large number of possible streamlines for every seed point. Therefore, you need to calculate a much larger number of streamlines using probabilistic tractography.
>>
>> It very much depends on your application if -number 100000 is enough. The easiest way to find out would be to do your processing all the way for one participant; then repeat the processing, say, 10 times, starting from streamtrack. Have a look at the metric you are interested in, and how variable it is across those 10 trials. If your results are fairly tight, 100000 is enough; if there is huge variation, you will need to increase the number of streamlines.
>>
>> Btw, FA initcutoff 0.6 and cutoff 0.3 seems fairly high to me. The MRtrix defaults are 0.2 (init) and 0.1; this works very well in my experience. I'm not sure what the cutoffs are for other software.
>>
>> Cheers
>> Kerstin
>>
>>
>>
>>
>> On 12/05/2011, at 12:39 AM, Luis Morís wrote:
>>
>>> Hi again,
>>>
>>> Thanks for your quick answer René.
>>>
>>> My data is 31 directions with b=1000. The point is that no matter how
>>> much fibers I put in the -number parameter, it always finds enough
>>> fibers even a couple millions fibres. Previously I was working with
>>> Medinria and I had around 30000 fibres because seeding was dependant
>>> on resolution, one seed per voxel.
>>>
>>> My parameters are FA cutoff 0.3 and initcutoff 0.6 min and max length
>>> 10 and 200 mm respectively, at the moment I'm doing a whole brain
>>> seeding. My problem is how to set this number of tracks. If I use a
>>> -number of let's say a hundred thousand, will I have a nice
>>> distribution of tracks that will be representantive of the brain's
>>> track population?
>>>
>>>
>>> Thanks in advance,
>>>
>>> Luis.
>>>
>>> On Wed, May 11, 2011 at 4:18 PM, René Besseling
>>> <r.m.h.besseling at gmail.com> wrote:
>>>> Hi Luis,
>>>>
>>>> Just type streamtrack -h; the tracks that are selected are those of
>>>> sufficient length (-minlength option) and sufficiently high FA or FOD
>>>> amplitude (-cutoff value). Note that the -maxnum option sets the
>>>> maximum allowed number of trials. If your data is of low quality, the
>>>> number of tracks generated can be lower than what you set by -number
>>>> if not enough valid tracks can be constructed within -maxnum trials.
>>>> Furthermore, tractography start points (seed points) are selected at
>>>> random within the ROI you specify using -seed, so if you use the same
>>>> streamtrack command twice, the result will be different (due to
>>>> different random seed points). Information about a streamtrack result
>>>> can be obtained by typing track_info your_file_name.tck
>>>>
>>>> Best regards,
>>>>
>>>> René
>>>>
>>>> On Wed, May 11, 2011 at 3:54 PM, Luis Morís
>>>> <luis.moris.fernandez at gmail.com> wrote:
>>>>> Hello everyone,
>>>>>
>>>>> I write because I have started to use mrtrix, and I've found a doubt.
>>>>> To be able to do tractography you have to use the parameter -number
>>>>> that will set the exact amount of tracks that will be calculated.
>>>>>
>>>>> My doubt is how does the program decide which tracks to include or
>>>>> not. If I set it to 1000 and I run it again with 1000000. What
>>>>> difference do I find between the two set of tracks? Is it purely
>>>>> random, depends on the length, the FA...
>>>>>
>>>>> Thanks a lot in advance,
>>>>>
>>>>> Luis.
>>>>> _______________________________________________
>>>>> Mrtrix-discussion mailing list
>>>>> Mrtrix-discussion at www.nitrc.org
>>>>> http://www.nitrc.org/mailman/listinfo/mrtrix-discussion
>>>>>
>>>>
>>> _______________________________________________
>>> Mrtrix-discussion mailing list
>>> Mrtrix-discussion at www.nitrc.org
>>> http://www.nitrc.org/mailman/listinfo/mrtrix-discussion
>>
>>
More information about the Mrtrix-discussion
mailing list