[MINC-users] running voxel-wise mixed-effects modeling on huge dataset with Python Rpy2 and pyezminc

[Vladimir S. Fonov]

Hello Everybody,

just wanted to share news about a new tool for running voxel-level 
statistical analysis on minc files, an alternative to RMINC and glim_image.

It is under active development, but it is already usable:

https://github.com/BIC-MNI/pyezminc/tree/develop

in short it is a python module to interface with MINC, mainly developed 
by Haz-Edine from NeuroRX with my additions. It is using cython (I 
recommend using the latest version 0.20 ) and numpy to interact with 
minc files using EZminc API. I have extended it to be able to stream 
data to and from multiple  minc files to easily perform voxel-level 
statistical analysis.
Integrating with Rpy2 ( R interface for Python) allows to easily run 
statistical tests.


For example: 
https://github.com/BIC-MNI/pyezminc/blob/develop/examples/glim_image.py 
- performs a simple voxel-level t-test on difference between two cohorts 
, using DBM data from 64 subjects (with ROI covering the whole brain) 
execution time is 14m56s , equivalent analysis performed using 
glim_image on the same computer takes 3m12s.


The advantage is it is possible to use more complicated statistical 
models (similar to RMINC), next example: 
https://github.com/BIC-MNI/pyezminc/blob/develop/examples/lme_image.py - 
performs voxel-level mixed-effect modeling (quite slowly).


Fortunately, it is pretty simple to parallelize execution of statistical 
modeling in python, example number 3 - running voxel-level mixed-effect 
modelling on a cluster using SCOOP ( https://code.google.com/p/scoop/ ) 
: 
https://github.com/BIC-MNI/pyezminc/blob/develop/examples/parallel_lme_image.py 



Running full-brain linear mixed effect model on 899 scans ( volume of 
mask is 235818 2mm^3 voxels) on a 16 core machine took approximately 23 
minutes. Note that the example doesn't try to play nice with the memory, 
so overall memory used was approximately 10Gb. It will be pretty trivial 
to significantly reduce the memory requirement by not submitting all the 
computation jobs at once , but in chunks.
It should be also possible to make scoop play nicely with SGE ( 
http://scoop.readthedocs.org/en/0.7/usage.html#startup-scripts-cluster-or-grid 
)


p.s. pyezminc depends on a modern installation of MINC (such as that 
found in minc-toolkit). And could be compiled and installed in your 
private python library by running "python setup.py install --user" . 
Currently dependencies are : Cython , numpy and scipy . Examples also 
use Rpy2 and scoop . All these libraries can be installed by running 
"python setup.py install --user" without the need to have root access on 
your system.

--
  Have Fun,

Vladimir S. Fonov