[pymvpa] PyMVPA for fiber bundle segmentation (PBC2009)

[Emanuele Olivetti]

Dear All,

I'm on holidays now after months of super intensive work. So let me
summarize my recent activities with PyMVPA and praise once again our
beloved toolbox.

In March 2009 the Pittsburgh Brain Competition started proposing
several tasks concerning the development of fully automated software
systems to analyze neuroimaging data (fMRI, DW-MRI):

http://www.braincompetition.org

I participated to the competition and worked on a joint method to
address both challenges 1 (cortex areas segmentation) and 2 (fiber
bundle segmentation) simultaneously, with the help of some of my
colleagues. Unfortunately I did not had enough time to work out full
solutions for both, so in the end I addressed only challenge 2: fiber
bundle segmentation. In short this task is about finding specific
fiber bundles (e.g. arcuate fasciculus) as a subset of fiber tracts of
the tractography in a subject starting from an example of that bundle
segmented by an expert on another subject's tractography.

Needless to say I developed my solution on top of PyMVPA, even though
surely not in the smartest way :D. An abstract of the method is here:

http://sfcweb.lrdc.pitt.edu/pbc/2009/sites/sfcweb.lrdc.pitt.edu.pbc.2009/files/abstract.txt

It seems that not many dared to participate to challenge 2 ;-)
Nevertheless our attempt got an honorable mention:

http://sfcweb.lrdc.pitt.edu/pbc/2009/?q=conference

which make me reasonably proud but should be a good point for PyMVPA
first. Without PyMVPA it wouldn't be possible at all to make that
effort in such a short time (~2 months). PyMVPA proved to be extremely
effective in implementing a complete new solution.

It was my first time with diffusion/tractography data but then got
very interested in the topic. In June/July I extended the proposed
solution to take into account both DW-MRI and fMRI (resting state)
data, i.e., using jointly both to segment automatically fiber bundles
(which was my true goal during the competition). The method is able to
do joint analysis of the data from the two modalities in a uniform way
(only very very little code is modality-specific).

This extra work, which falls into the field of joint analysis of
structural and functional data, became an article - currently
submitted to a journal. Let's hope reviewers will like uncommon
machine learning ideas :-D

Maybe someone is wondering about the code I wrote. Yes, it is messy
and long. But since my solution seems quite flexible I expect to work
on it for quite a bit in future :) . So I hope to release something
soon.


Time to go to swim now!


Thanks again to all PyMVPA people.


Best,


Emanuele