[pymvpa] the effect of ROI size on classification accuracy

Meng Liang meng.liang at hotmail.co.uk
Fri Jul 18 17:06:12 UTC 2014

Dear Jo,
Thank you so much for your quick reply! 
The example in your blog showing the discontinuous information detection is very interesting to me. Although I also thought that classification accuracy could drop by adding more voxels in the ROI especially if those voxels are just noise or less informative so they could, in theory, dramatically change the distribution of the data points in the hyperspace and thus change the classification accuracy dramatically. But I didn't expect by adding equally informative voxels could also dramatically change the classification performance. Nice to know!
Regarding the 5-mm Gaussian smoothing, I was actually testing the effect of a range of smoothing parameters on the classification accuarcy, namely, 0-mm, 5-mm, 10-mm and 15-mm. My original voxel size was 3x3x3mm and was then resampled to 2x2x2mm when normalized to MNI space. I didn't really expect to see a big drop in classification accuracy when changing the data from 0-mm smoothing to 5-mm smoothing, mainly because this is a between-subject MVPA (5-mm smoothing could/should even improve the performance by removing some inter-subject variability in brain anatomy). That's one reason I'm puzzled about the results. Having said that, sigma=5mm smoothing equals FWHM=11.8mm smoothing, so the smoothed image does look considerably smoother than the unsmoothed image. 
I was also wondering whether this was due to some mistakes. But all results were generated from the same code (the only difference is the nifti image files being read into the script). Not sure what other things to check... Ideas?

> Date: Fri, 18 Jul 2014 10:03:15 -0500
> From: jetzel at artsci.wustl.edu
> To: pkg-exppsy-pymvpa at lists.alioth.debian.org
> Subject: Re: [pymvpa] the effect of ROI size on classification accuracy
> First, 30432 is a LOT of voxels for a ROI-based analysis. Linear svm can 
> "pool" information from many voxels and have discontinuous information 
> detection 
> (http://mvpa.blogspot.com/2013/09/linear-svm-behavior-discontinuous.html), 
> all of which makes it hard to interpret results when two ROIs are 
> massively different sizes.
> Also, 5 mm Gaussian smoothing isn't all that much at most voxel sizes; 
> were your voxels very small? If the voxels were 3x3x3 mm or so I 
> wouldn't expect a bit of smoothing to make much difference.
> Regarding the actual accuracies you report, I find the dataset #1 ones 
> fairly plausible; I'd interpret as ROI A informative, ROI B informative 
> (surprisingly so, given how big it is), and ROI A + B informative. 
> Probably the three accuracies are not significantly different; there are 
> enough informative voxels in both ROIs to have good classification, and 
> adding A to B just adds redundant information (or doesn't add enough 
> uniquely-informative voxels to improve accuracy further).
> But, combined with the accuracies for dataset #2, things are stranger. 
> My main suggestion is to double-check everything: the most likely reason 
> I can think of for such a big and inconsistent influence of smoothing is 
> that something went wrong. For example, visually confirm that the 
> smoothed images look like they have the proper amount of smoothing, and 
> run both sets of images through the same code.
> good luck,
> Jo
> On 7/18/2014 9:38 AM, Meng Liang wrote:
> > Dear experts,
> >
> > I got some weird results when running MVPA (to distinguish two different
> > stimulus categories) on two fMRI datasets with different smoothing using
> > three different ROIs. I would like to know your opinion on why this
> > could happen.
> >
> > I used linear SVM.
> > The two datasets are from the same data acquisition but with different
> > spatial smoothing: (1) without any spatial smoothing and (2) Gaussian
> > smoothing with sigma=5mm
> > Three ROIs: (A) brain area A containing 357 voxels, (B) brain area B
> > containing 30432 voxels, and (C) brain areas A+B containing
> > 357+30432=30789 voxels.
> >
> > The classification accuracies when using dataset#1:
> >        0.750 for A,
> >        0.792 for B,
> >        0.792 for A+B
> >
> > The classification accuracy when using dataset#2:
> >        0.875 for A,
> >        0.667 for B,
> >        0.583 for A+B
> >
> > so, using unsmoothed data, combining A and B did not change the
> > classification accuracy. However, using smoothed data, combining A and B
> > reduced the classification accuracy considerably and the accuracy was
> > not significantly higher than chance level any more (all other
> > accuracies were significantly higher than chance level according to
> > permutation test).
> >
> > I would be grateful if anyone could let me know your thoughts why
> > changing the ROI size has different effect on smoothed and unsmoothed data.
> >
> > Best,
> > Meng
> >
> >
> > _______________________________________________
> > Pkg-ExpPsy-PyMVPA mailing list
> > Pkg-ExpPsy-PyMVPA at lists.alioth.debian.org
> > http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/pkg-exppsy-pymvpa
> >
> -- 
> Joset A. Etzel, Ph.D.
> Research Analyst
> Cognitive Control & Psychopathology Lab
> Washington University in St. Louis
> http://mvpa.blogspot.com/
> _______________________________________________
> Pkg-ExpPsy-PyMVPA mailing list
> Pkg-ExpPsy-PyMVPA at lists.alioth.debian.org
> http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/pkg-exppsy-pymvpa
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.alioth.debian.org/pipermail/pkg-exppsy-pymvpa/attachments/20140718/796c3ff8/attachment.html>

More information about the Pkg-ExpPsy-PyMVPA mailing list