Those data arenıt so high quality (long TR, low resolution, short runs) so
perhaps thatıs why it isnıt working well (FIX works near 100% on TR=720ms,
2x2x2mm resolution, and 15 minute runs). With worse quality data, the ICA
will have more difficulty splitting signal and noise. It is also
possible that your manual training is not very good (if you havenıt
segregated signal and noise components properly).
Peace,
Matt.
On 1/22/16, 6:31 AM, "FSL - FMRIB's Software Library on behalf of Julia
Schumacher (PGR)" <[log in to unmask] on behalf of
[log in to unmask]> wrote:
>There is one run of resting state data per subject, with TR=3000ms, 128
>volumes, resolution: 2x2x6mm.
>
>Cheers,
>Julia
>
>________________________________________
>Von: FSL - FMRIB's Software Library <[log in to unmask]> im Auftrag von
>Matt Glasser <[log in to unmask]>
>Gesendet: Freitag, 22. Januar 2016 11:43
>An: [log in to unmask]
>Betreff: Re: [FSL] low TNR in FIX training
>
>What kind of data is this (resolution, TR, number of minutes, number of
>runs)?
>
>Peace,
>
>Matt.
>
>On 1/22/16, 4:28 AM, "FSL - FMRIB's Software Library on behalf of Julia
>Schumacher" <[log in to unmask] on behalf of [log in to unmask]>
>wrote:
>
>>Dear FSL experts,
>>
>>I would like to use FIX to clean my data and have tried to train it on 47
>>hand-labelled subjects. Everything seems to have worked fine, however,
>>the TNR is very low, even for higher thresholds.
>>This is the mean TPR and TNR from the LOO-classification for the
>>different thresholds:
>>
>>mean
>>99.8 99.4 97.9 95.8 91.9 88.6 83.7 77.4
>>5.0 8.7 14.4 22.0 32.6 40.3 47.7 55.1
>>
>>Does anyone have an idea what could be the problem? It might just be due
>>to my data or the hand labels that I provided, but I was wondering if
>>there might be some general problem or some general way to increase the
>>classification accuracy.
>>
>>Any help would be greatly appreciated!
>>Thank you,
>>Julia
|