Hi all,
I mentioned before that I was at a conference
meeting (Organization of Human Brain Mapping, 2015 http://ohbm.loni.usc.edu/)
where I had the great chance to meet with my mentors. Now, it's time to update
on what was done during those days and during the week after (last week).
As stated in my proposal, the project consists of
classifying a brain T1 MRI into “tissue classes” and estimating the partial
volume at the boundary between those tissues. Consequently, this is a brain
segmentation problem. We decided to use a segmentation method based on Markov
Random Field modeling, specifically the Maximum a Posteriori MRF approach
(MAP-MRF). The implementation of a MAP-MRF estimation for brain tissue
segmentation is based on the Expectation Maximization (EM) algorithm, as
described in Zhang et al. 2001 ("Segmentation of brain MR images through a
hidden Markov random field model and the expectation-maximization
algorithm," Medical Imaging, IEEE Transactions on, vol.20, no.1, pp.45,57,
Jan 2001). The maximization step is performed using the Iterative Conditional
Modes (ICM) algorithm. Thus, together with my mentors, we decided to first work
on the ICM algorithm. I started working on it during the Hackathon at OHBM and
finished it up last week. It is working now and I already shared it publicly to
the rest of the DIPY team. I submitted my first pull request called:
WIP: Tissue classification using MAP-MRF
https://github.com/nipy/dipy/pull/670#partial-pull-merging
There was a lot of feedback from all the team,
especially regarding how to make it faster. The plan for this week is to
include the EM on top of the ICM and provide the first Partial Volume
Estimates. Will do some testing and validation of the method to see how it
performs compared to other publicly available methods such as FAST from FSL
(http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FAST).
