It would be lovely if FSL incorporated multi-threading for melodicıs PCA.
Did you concatenate the data yourself outside of melodic or use melodicıs
text file input? MIGP will reduce unstructured noise in the data
somewhat, which might make the auto dimensionality selection choose to
find more components. It wold probably help if you provided the melodic
command lines you used for your tests.
Peace,
Matt.
On 8/9/15, 12:08 AM, "FSL - FMRIB's Software Library on behalf of Parnesh
Raniga" <[log in to unmask] on behalf of [log in to unmask]> wrote:
>We are planning on using the MIGP option to perform group based analysis
>on a large dataset. Before running the full analysis, we tried it on a
>subset of 50 subjects and compared to temporal concatenation with
>estimation of number of components performed by melodic.
>
>We find that the number of components estimated using MIGP to be
>consistently 3-4x times the number using temporal concat (13,15,19) vs
>(63,69). The spatial maps however are consistent.
>
>Have others using MIGP noticed any differences between temporal concat
>and MIGP?
>
>On a side note, we also have modified the melodic code (not part of the
>above test) to use multithreaded blas for the PCA. This significantly
>speeds up analysis. I am happy to provide a patch for anyone interested
>once we have done some more testing.
|