[Mrtrix-discussion] statistical analysis after fiber tracking

[Thijs Dhollander]

Hi Zhuang (& others),

In the mean time, the ISMRM program became public; the abstract I was talking about before, can be found in the "Fibers & Tractography" session (at http://www.ismrm.org/13/session77.htm ).  The abstracts are only available to (early) registrants at the moment though.  Zhuang: I will send you a copy.
On a slightly unrelated note: I can certainly advise this whole scientific session to anyone attending ISMRM; there's a lot of exciting and really interesting stuff in there! (by this, of course, I also shamelessly promoted my own talk ;-) ).

There certainly is a lot of value in fiber tracking, even as part of quantificiation - if done right.  The more recently proposed TWI is a big step in the right direction - if again done right.  I'm mostly on the sceptical side - or at least on the rather careful side - because there are so many ways to apply these methods: for every "right" way to do it, there's an increasing amount of ways to do it "wrong" as these methods become more complex and dependent on a lot of parameters and user made choices.

Yet another interesting contribution to quantification is found in another abstract in the "Cortex, Connections & Connectomes" session ( at http://www.ismrm.org/13/session83.htm ): "Tractographic Threshold-Free Cluster Enhancement: Whole-Brain Statistical Analysis of Diffusion MRI Measures in the Presence of Crossing Fibres" by David.  It's a great example how tractography can be cast in the role it is most suited for: a "segmentation"-like role; in this case for determining spatio-angular neighbourhoods.  The final measure used for quantification is still AFD which is related much more closely to the actual measured data!

Cheers,
Thijs

From: Zhuang Song [mailto:zhuang.song at gmail.com]
Sent: Wednesday, March 27, 2013 00:28
To: Thijs Dhollander
Subject: Re: [Mrtrix-discussion] statistical analysis after fiber tracking

Hi Thijs,

Thank you so much for your thoughtful responses. It took me some days to think and try things that you suggested. The papers that you recommended are also very informative and relevant, especially the AFD paper.

Thanks for letting me know the 'colour' option of the tracks2prob command. I am interested to separate different bundles and only analyze the fibers in certain directions. Since the direction of the fiber of interest is well known, it seems possible to eliminate the ambiguity of the color coding in the tracks2prob command. There are some technical problems to derive directional information from this command. Since those problem are related to the implementation of MRtrix, I will post in their maillist later.

It is very interesting that you use directional information for different purposes. Unfortunately I won't go to ISMRM this year. But I will definitely appreciate if you could let me know when your journal paper comes out. Do you obtain directional information from the output of MRtrix? Would you mind to describe a bit how you did it?

I strongly agree with you that fiber tracking is only for segmentation and visualization but not for quantification. I appreciate your cautious notes. Indeed there are a lot of information that can be derived from the ODF, such as AFD, that can be very powerful to detect subtle changes in white matter tracts. Thank you again for sharing your thoughts, which are very helpful for me.


Best regards,
Zhuang

---------------------------------------------------------------------
Zhuang Song, Ph.D.
Associate Research Scientist
Center for Imaging of Neurodegenerative Diseases
Department of Radiology and Biomedical Imaging
University of California, San Francisco
Phone: 415-221-4810 x6454
Fax: 415-668-2864


On Fri, Mar 22, 2013 at 3:12 AM, Thijs Dhollander <thijs.dhollander at uzleuven.be<mailto:thijs.dhollander at uzleuven.be>> wrote:
Hi Zhuang

Here's some thoughts about your questions:


1)      At the moment, there's the "colour" option for the tracks2prob command.  However, one colour from that map can in general still correspond to a maximum of 4 different directions.  As a simple example, consider a yellow colour: it could be a direction from front-left to back-right, but also from front-right to back-left (both mix up red and green, resulting in yellow; it's the same "problem" also present in a DEC FA map).  So, if you really need the true direction, the colours are not specific enough.  Apart from that, I'm not exactly sure what you mean by "where crossing fibers have been removed": do you want everything mixed up (i.e. an average direction, ignoring the fact that there are possible contributions from different crossing bundles), or rather separated (i.e. so you can consider the bundles separately without influence from each other).  I'm presenting an abstract at the upcoming ISMRM in Salt Lake City that includes a technique which can do both of these for you (however, in the abstract, I'm subsequently using this kind of information for an entirely different purpose than statistical analysis).  A journal paper with more details is on its way as well.  I've extended the tracks2prob command's functionality to be able to calculate it automatically (still version 2.10, so I might still merge it with the changes in 2.11); I might release that code soon (when the journal paper is more or less out in the open).

2)      The best option here is to consider the FOD's magnitude, which is related to the apparent fiber density (AFD) metric.  There's a hefty paper on that (with a lot of indept discussion): http://www.sciencedirect.com/science/article/pii/S1053811911012092 .  It might probably be available in the next major release of MRtrix.  A possibility that is "accessible" in the mean time, is to consider the average AFD (over all directions) in each voxel.  To generate it... well you don't even have to, cause it's already present in your FOD file resulting from CSD: just open it directly in the MRtrix viewer (not in the orientation plot part, but just directly as an image).  The first thing you'll see, is the order 0 SH coefficient map from the FOD's, which is -up to a constant factor- directly showing you an average AFD map.
Appart from that, there also exists something called the "GFA" (generalized FA), which basically is the standard deviation normalized by the root mean square of the FOD's (or any other angular function) magnitude.  It's not very "mainstream", and suffers more or less (maybe a bit less) from the same problems as the FA.  Not very worthwhile to put effort in, in my opinion.

3)      That's a tough one (feel the controversy coming up... ;-)).  In general, it fully depends on your fiber tracking algorithm, strategy, parameters, and just about everything else involved (e.g. seeding strategy/distribution, amongst others...).  That basically renders it very tricky for a true physical/biological/...real-worldish.../... interpretation.  I've been ranting a bit on that on different occasions last year (best source is probably last year's ISMRM abstract I attached to this mail).  But being critical for my own work on that: it just considered 1 kind of way to do TDI (more or less the most basic one, based on how it's done in the original TDI paper).  Everything changes with your specific tracking setup.  Here's another article that discusses some things on the (lack of) meaning of track counts (also applies to fractions, they deliver the same contrast): http://www.sciencedirect.com/science/article/pii/S1053811912007306  .  As you might notice, I'm (a bit) on the sceptical side here.  It could be done better, if you do the tracking in a "responsible" way though.  Responsible in this context, would in my opinion be the combination of these strategies: http://www.sciencedirect.com/science/article/pii/S1053811912005824  and  http://www.sciencedirect.com/science/article/pii/S1053811912011615 .  They should also be included in the next major release of MRtrix!  Whatever you do, you have to carefully think about what it still means: don't forget that tracks are just generated by an algorithm with a lot of heuristic assumptions.  Fiber tracking may as such be a good (great) segmentation and visualisation tool, the densities on the other hand where never (originally) ment to be something really meaningful.

Hope this helps; I'm sure that there are a lot of other people around here that also have interesting ideas and opinions on these (difficult) topics!

Regards,
Thijs


From: mrtrix-discussion-bounces at public.nitrc.org<mailto:mrtrix-discussion-bounces at public.nitrc.org> [mailto:mrtrix-discussion-bounces at public.nitrc.org<mailto:mrtrix-discussion-bounces at public.nitrc.org>] On Behalf Of Zhuang Song
Sent: Thursday, March 21, 2013 19:09
To: mrtrix-discussion at public.nitrc.org<mailto:mrtrix-discussion at public.nitrc.org>
Subject: [Mrtrix-discussion] statistical analysis after fiber tracking

Dear MRtrixers,

There are several questions regrading statistical analysis after fiber tracking, which may not necessarily relate to each other.

1) Given the tracking output file *.tck, is there any way to read out the mean direction of selected fibers in every voxel where crossing fibers have been removed? This voxelwise directional information is anatomically specific and therefore can be very useful in statistical analysis.

2) For the purpose of statistics (not fiber tracking), what kinds of metric would you recommend to derive from either the original ODF or fiber ODF? I haven't seen much this kind of work yet in literature. It would be nice to have some brainstorming here.

3)  I was wondering how to interpret properly the  physical meaning of the track number or density that streamtrack generates. The program streamtrack can generate arbitrary large number of tracks by manipulating the option "-number". The track density (i.e. "-fraction" outputs of tracks2prob) seems to remain quite stable even if I increase the total tracking number 1000 times in streamtrack with option "-number". Different than the TDI method based on whole brain tracking, I am only interested in the tracks within a ROI. It seems to me it makes more sense to use the track density instead of track numbers in statistical analysis. Is it reasonable to do so even if the tracking is restricted in a ROI?

Thanks,
Zhuang





-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.nitrc.org/pipermail/mrtrix-discussion/attachments/20130408/d645d15a/attachment-0001.html