Back to the main page.
Bug 2437 - consider using image opening to post-process thresholded map when doing (source) statistics
| Status | ASSIGNED |
| Reported | 2014-01-16 15:15:00 +0100 |
| Modified | 2014-01-22 13:49:11 +0100 |
| Product: | FieldTrip |
| Component: | core |
| Version: | unspecified |
| Hardware: | PC |
| Operating System: | Windows |
| Importance: | P5 enhancement |
| Assigned to: | Eelke Spaak |
| URL: | |
| Tags: | |
| Depends on: | |
| Blocks: | |
| See also: |
Eelke Spaak - 2014-01-16 15:15:12 +0100
Created attachment 582 opening demo There is a thresholded n-dimensional binary map involved in all clustering routines, derived from comparing the computed statistic to the specified critical threshold. For sensor-level time/frequency/channel clustering we have the parameter cfg.minnbchan, which removes samples from this binary map in case the time/frequency clusters are connected by a number of channels less than the value of cfg.minnbchan. This makes sense because from physiological prior info we can expect 'real' clusters to be connected by more than just a narrow bridge of samples. I recall several posts on the mailing list from people asking about a similar cfg.minnbtim or cfg.minnbfreq parameter, separating clusters that are connected by a narrow bridge in the time or freq dimension. This would also make sense, I think. However, this is not possible to implement in the same way as cfg.minnbchan, as throughout findcluster() the time/freq dimension is treated as a regular 2D matrix (which is good, because then labelling of the clusters is very fast through spm_bwlabel). Similarly, I just ran into this issue when doing source statistics: I observed a big cluster consisting of two physiologically plausible 'blobs' connected by a physiologically highly implausible narrow bridge of samples (in 3D the probability of getting such a wacky bridge is of course much higher than in 2D). The clustering routine of course treated this cluster as one, making the results hard to interpret. After a short amount of Googling I found some stuff about morphological binary image processing. This is a useful link: http://www.johnloomis.org/ece563/notes/BinaryImages/morph/morph.html and this one is specific to Matlab: http://www.mathworks.nl/help/images/morphology-fundamentals-dilation-and-erosion.html . The process of 'opening' (an erosion followed by a dilation) seems to do exactly what I would want here: destroy narrow bridges, retain only major features of the image. See the attachment for a demonstration of what it does on a 2D thing I drew: subplot(2,1,1); imagesc(im); colormap gray; title('original'); subplot(2,1,2); imagesc(imopen(im, ones(3,3))); colormap gray; title('opened'); I think it would be a good idea to add this functionality to clusterstat. At least first we might implement it in the case of source statistics, as it is very straightforward there, just add a call to imopen() before doing the labelling of clusters. For time/freq clustering I'm not sure how this would combine with the handling of neighbouring channels (although it might be very simple in that case, too). Does anyone see a problem with me implementing this? Maybe we could discuss how to proceed during the meeting.
Robert Oostenveld - 2014-01-16 18:28:07 +0100
From the top of my head: I think there is a minnbchan option somewhere to prevent narrow bridges (in channel-dimension?) imopen/imclose/imerode/imdilate all are related to this. There is also something with a "structural element" that one can specify for some of the image toolbox operations. I suggest to look into how it is presently done for channels (somewhere in the code of Eric), and make it consistent.