help
489 Subscribers 
help > One large cluster in Seed-Voxel analysis
Showing 1-3 of 3 posts
Nov 8, 2019 06:11 AM | Islay Davies - The University of Queensland
One large cluster in Seed-Voxel analysis
Hi
I'm running a seed-to-voxel functional connectivity analyses using the Conn toolbox on a sample of young adolescents (n = 340) acquired on a 3T Prisma (2 x 5 minute scans per participant, distortion correction prior to Conn toolbox with FSL topup).
I can't see any obvious errors with my pre-processing, denoising, and first level analyses; however, my second-level analyses results don't look correct. Basically, whichever atlas seed region I select as a source (using the default atlas in Conn) results in one large cluster of connected voxels (both positive & negative connections). The pattern of voxel connections do change depending on the seed, but it is seemingly always one large cluster containing nearly all voxels.
Any ideas where I've gone wrong? I haven't deviated too far from any of the default settings (just small changes to smoothing & the number of denoising parameters).
Kind Regards
Islay
I'm running a seed-to-voxel functional connectivity analyses using the Conn toolbox on a sample of young adolescents (n = 340) acquired on a 3T Prisma (2 x 5 minute scans per participant, distortion correction prior to Conn toolbox with FSL topup).
I can't see any obvious errors with my pre-processing, denoising, and first level analyses; however, my second-level analyses results don't look correct. Basically, whichever atlas seed region I select as a source (using the default atlas in Conn) results in one large cluster of connected voxels (both positive & negative connections). The pattern of voxel connections do change depending on the seed, but it is seemingly always one large cluster containing nearly all voxels.
Any ideas where I've gone wrong? I haven't deviated too far from any of the default settings (just small changes to smoothing & the number of denoising parameters).
Kind Regards
Islay
Nov 12, 2019 01:11 PM | Alfonso Nieto-Castanon - Boston University
RE: One large cluster in Seed-Voxel analysis
Hi Islay,
That sounds perfectly fine. With a sample of 340 subjects you probably have reasonable power to detect (using the default height threshold of p<.001) discrepancies in Fisher-transformed correlation coefficients of around dZ = ±0.05, which, if we are just estimating raw average connectivity values across all subjects, can be expected to cover most of the brain. As long as the patterns of positive/negative connectivity looks reasonable (e.g. like the example attached for MPFC seed; this is from a sample of 198 subjects only but it should still look similar to your case) the fact that you get significant effects in most brain voxels is perfectly fine (it only speaks about the relatively high sensitivity/power of the analysis to detect even small discrepancies away from the -admittedly somewhat uninformative- zero-connectivity null hypothesis of the one sample t-test being performed here). If you prefer to see more spatially-informative patterns in these analyses simply use more conservative height thresholds (e.g. use T>10 height threshold; in the "results explorer" view switch the "p-uncorrected" option to "T/F/X stats") and/or switch to one-sided stats.
Best
Alfonso
Originally posted by Islay Davies:
That sounds perfectly fine. With a sample of 340 subjects you probably have reasonable power to detect (using the default height threshold of p<.001) discrepancies in Fisher-transformed correlation coefficients of around dZ = ±0.05, which, if we are just estimating raw average connectivity values across all subjects, can be expected to cover most of the brain. As long as the patterns of positive/negative connectivity looks reasonable (e.g. like the example attached for MPFC seed; this is from a sample of 198 subjects only but it should still look similar to your case) the fact that you get significant effects in most brain voxels is perfectly fine (it only speaks about the relatively high sensitivity/power of the analysis to detect even small discrepancies away from the -admittedly somewhat uninformative- zero-connectivity null hypothesis of the one sample t-test being performed here). If you prefer to see more spatially-informative patterns in these analyses simply use more conservative height thresholds (e.g. use T>10 height threshold; in the "results explorer" view switch the "p-uncorrected" option to "T/F/X stats") and/or switch to one-sided stats.
Best
Alfonso
Originally posted by Islay Davies:
Hi
I'm running a seed-to-voxel functional connectivity analyses using the Conn toolbox on a sample of young adolescents (n = 340) acquired on a 3T Prisma (2 x 5 minute scans per participant, distortion correction prior to Conn toolbox with FSL topup).
I can't see any obvious errors with my pre-processing, denoising, and first level analyses; however, my second-level analyses results don't look correct. Basically, whichever atlas seed region I select as a source (using the default atlas in Conn) results in one large cluster of connected voxels (both positive & negative connections). The pattern of voxel connections do change depending on the seed, but it is seemingly always one large cluster containing nearly all voxels.
Any ideas where I've gone wrong? I haven't deviated too far from any of the default settings (just small changes to smoothing & the number of denoising parameters).
Kind Regards
Islay
I'm running a seed-to-voxel functional connectivity analyses using the Conn toolbox on a sample of young adolescents (n = 340) acquired on a 3T Prisma (2 x 5 minute scans per participant, distortion correction prior to Conn toolbox with FSL topup).
I can't see any obvious errors with my pre-processing, denoising, and first level analyses; however, my second-level analyses results don't look correct. Basically, whichever atlas seed region I select as a source (using the default atlas in Conn) results in one large cluster of connected voxels (both positive & negative connections). The pattern of voxel connections do change depending on the seed, but it is seemingly always one large cluster containing nearly all voxels.
Any ideas where I've gone wrong? I haven't deviated too far from any of the default settings (just small changes to smoothing & the number of denoising parameters).
Kind Regards
Islay
Nov 18, 2019 08:11 AM | Islay Davies - The University of Queensland
RE: One large cluster in Seed-Voxel analysis
Thank you Alfonso,
Kind Regards
Islay
Kind Regards
Islay