Hello,
TFCE in randomise is effectively providing voxelwise, FWE-corrected p-values. If you cluster this output at ( say a 1-p of 0.95 ), then for a reported cluster, you could say that a cluster of connected voxels was found above that significance threshold. This is not quite the same thing as the -c option, which will assign a single p-value based on the null-distribution of cluster-sizes, but carries something of the same intent.
Kind Regards
Matthew
> Hi Matthew,
>
> I’m interested in knowing whether TFCE can provide a single p-value for a cluster. Is there a way to interpret the output corrp files from TFCE in order to provide such a measure? From what you say below, it’s not simply the case that -c guides TFCE to provide cluster level p-value information.
>
> Perhaps I am missing the point of TFCE i what I say! Is the only measure from TFCE the cluster extent (number of voxels) and the peak voxel p-vlaue within the cluster? If so, how does one interpret how significant the cluster actually is without a single p-value?
>
> Cheers,
> Donal
>
>
> On 7 Dec 2015, at 16:25, Matthew Webster <[log in to unmask]> wrote:
>
>> Hello Donal,
>> The -c and -T options work independently on the same "raw" input statistic. You can use them together or separately in an analysis, although you need to correct for FWE if you intend to use both outputs when analysing your results. Note that TFCE has been found to be more sensitive than standard cluster-based thresholding.
>>
>> Kind Regards
>> Matthew
>>
>>> Hello Matthew,
>>>
>>> Many thanks for your response! Are the -c and -T (TFCE) options mutually exclusive, by which I mean do they run different algorithms to determine clusters? Or is it the case that -c is an option that alters the output of TFCE from voxelwise p-values to cluster single p-values?
>>>
>>> Cheers,
>>> Donal
>>>
>>> On 7 Dec 2015, at 15:18, Matthew Webster <[log in to unmask]> wrote:
>>>
>>>> Hello,
>>>> Randomise should already output p-value ( actually 1-p ) images for TFCE. You could threshold these at 0.95 with cluster to find discrete "clusters" of significant voxels. However, note that it is not appropriate to use the cluster's GRF options with TFCE output, as TFCE values do not have a known distribution ( randomise is calculating an empirical null ). If you specifically want to obtain clusters with a single p-value, then you should the -c or -C options in randomise, which make fewer assumptions than the GRF options in cluster.
>>>>
>>>> Hope this helps,
>>>> Matthew
>>>>
>>>>> Dear experts,
>>>>>
>>>>> I have run randomise using the TFCE option (-T), producing t-statistic maps. I am using FSL Version 5.0.9. I have subsequently run “cluster" on the randomise output in order to extract significant clusters, but I would like to retrieve the p-value of the clusters using the GRF option.
>>>>>
>>>>> I know that I need to supply --dlh (smoothness estimate), --volume (number of voxels in mask) and --pthresh (p-value threshold for clusters), where I use the “smoothest" function to determine values for --dlh and --volume.
>>>>>
>>>>> The “cluster” command I run is:
>>>>>
>>>>> cluster --in=output_tfce_thresh_tstat2.nii.gz --dlh=0.0239923 --volume=611340 --pthresh=0.95 -t 0.95 --oindex=output_cluster_index.nii.gz --olmax=output_lmax.txt --osize=output_cluster_size.nii.gz
>>>>>
>>>>>
>>>>> Unfortunately, I get “nan” returned for the cluster p-vlaue:
>>>>>
>>>>> Cluster Index Voxels P -log10(P) MAX MAX X (vox) MAX Y (vox) MAX Z (vox) COG X (vox) COG Y (vox) COG Z (vox)
>>>>> 1 1786 nan -nan 17.7 43 86 43 43.2 84 48.2
>>>>>
>>>>> whereas when I run without any GRF options, I see the following:
>>>>>
>>>>> Cluster Index Voxels MAX MAX X (vox) MAX Y (vox) MAX Z (vox) COG X (vox) COG Y (vox) COG Z (vox)
>>>>> 1 1786 17.7 43 86 43 43.2 84 48.2
>>>>>
>>>>>
>>>>> I am unsure why I am getting nan values when using GRF options. Any help would be much appreciated!
>>>>>
>>>>> Cheers,
>>>>> Donal
|