 |
 |
Hi Anderson, Thank you for the detailed response. If I understand correctly what you write: 2D clustering should work, there are some additional distortions and problems that will (slightly?) affect the p-values of the clusters but, in principle, the pipeline should not be totally/fundamentally wrong. Good. It could be that in our case there is slightly less distortion due to 2D projection as we are not imaging the whole head (so that there will be less clusters at the meridian) and rotate our ROI to the middle of the surface. Yes, please, it would very be interesting to try your functions for permutation test. Perhaps, at some point, those functions will be integrated in fsl (options in the command line tools)? -- Teemu Anderson M. Winkler kirjoitti 22.8.2012 kello 5.43: > Hi Teemu, > > This should reduce to the case of a NIFTI file with just one slice per subject, and I believe this should be handled correctly. Others could correct me if I'm wrong. > > However, I perceive other potential problems. Even though you are being very careful and using an equal-area projection to the plane, the spherical transformation (white->sphere) isn't an area-preserving process, such that different parts of the brain are differently scaled when the sphere is generated, and this will be propagated to the flattened map, and ultimately impact the results for cluster size. > > A second problem is that for the Mollweide and similar projections it's necessary to cut the surface. If, unluckily, a cluster lies at the meridian 180°, it will be split, influencing the inference. For this specific problem, perhaps a solution would be to use an azimuthal projection, placing the antipode of its centre in the medial aspect of the brain. A projection with this feature is the Lambert's, which is also an equal-area: http://en.wikipedia.org/wiki/Lambert_azimuthal_equal-area_projection > > A third problem is that both the spherical transformation and the projection to 2D add distortion. This creates more challenges as the non-stationarity becomes extremely pronounced across the map and residuals. And of course, a fourth problem is the extra interpolation step from the mesh to the regular voxel grid. > > Despite these, it's nice to hear that it seems to be working, with results as expected. Nonetheless, a possible solution for all these problems at once would be to do non-parametric inference (a permutation test) directly in the surface space. I have some functions in Matlab/Octave for facewise analysis that could be easily modified for vertexwise, and which I'm happy to share. If you think this could be of some help, let me know. > > All the best, > > Anderson > > > 2012/8/21 Teemu Rinne <[log in to unmask]> > Dear FSL Experts, > > We would like to do as follows: > (1) standard FEAT analysis in 3D native (EPI) space > (2) anatomical normalization in spherical standard space using Freesurfer > (3) anatomically normalized surfaces are projected (using equal area Mollweide projection) to a 2D space separately for each hemi > (4) the results of 3D FEAT analysis are transformed to 2D space for across the subjects analysis > > Finally, we would like to run FLAME (analysis across subjects) with cluster correction on the flattened 2D data. > > (In the flattened 2D space, EPI data files are similar to single slice data) > > We would like to verify whether the FSL cluster correction works ok in the 2D data. Our (practical) tests suggest that it does: the results of a full 3D analysis (feat, group analysis and cluster correction in 3D) flattened to 2D and the results of 2D group analysis (feat in 3D, group analysis and cluster correction in 2D) produce nearly identical results (some tiny clusters do not survive in the 2D analysis). > > Are there any noteworthy issues or potential pitfalls with this approach (cluster correction in 2D)? > > (We could live with the facts that some activations are lost in the 3D->2D projection, that some tiny but real clusters will not survive cluster correction in 2D, and that p-values associated with 2D clusters could be different than those of the corresponding 3D clusters) > > We would like to do the group analysis in the 2D space because, according to our understanding, it is not currently possible to use FSL cluster correction in the spherical space. We would like to use FSL tools throughout the statistical analysis (and not to use FLAME and, say, FS cluster correction in the spherical space). > > > Cheers, > > Teemu Rinne > Institute of Behavioural Sciences > University of Helsinki >