Hello,
         The paper: Group-PCA for very large fMRI datasets by S. Smith et al. discusses how melodic’s dimensionality reduction works. Melodic outputs spatially independent components.

Kind Regards
Matthew

--------------------------------
Dr Matthew Webster
FMRIB Centre
John Radcliffe Hospital
University of Oxford

> On 23 May 2018, at 14:53, Dani Hak <[log in to unmask]> wrote:
>
> Thanks for clarifying Prof. Smith and Matthew,
>
> I was actually wondering why the times series became shorter at the group ICA.
> I started from series of ~1500 points (500 for each of 3 subjects) and ended up with ~1000 (at the group level, without the dual regression step).
> I was wondering why the algorithm reduces the time series length, and if it assumes all the ICA components are independent in time.
>
> Best,
> Dani. 
>   
>
> On Wed, May 23, 2018 at 2:46 AM, Prof Stephen Smith <[log in to unmask]> wrote:
> Hi - the group-ICA doesn't know about subject-level timeseries by the time it has done the group-PCA dimensionality reduction - that's why Matthew is saying you need to then run dual-regression and look at single-subject (or averaged across subjects) temporal power spectra to help decide which group-level components are good/bad (though you can often tell just from the group spatial maps).  Your end goal sounds sensible though.
>
> Cheers.
>
>
>> On 22 May 2018, at 19:11, Dani Hak <[log in to unmask]> wrote:
>>
>> Thank's Matthew,
>>
>> Originally I wanted to use ICA to identify regions that are activated across my subjects population.
>> I was hoping to get these components from group ICA and be able to use the time series and spectral plots to differentiate noise from the signal.
>> If I understand you, this can to be done on the subject level only.
>> Is that correct?
>>
>> On Fri, May 18, 2018 at 6:40 AM, Matthew Webster <[log in to unmask]> wrote:
>>  Hi Dani,
>> These are the internal representations of the timecourses in MIGP mode. If you want the associated time courses you need to run dual regression of the spatial maps. This will generate least-squares time courses per subject as part of dr_stage1.
>>
>> Kind Regards
>> Matthew
>> --------------------------------
>> Dr Matthew Webster
>> FMRIB Centre
>> John Radcliffe Hospital
>> University of Oxford
>>
>>> On 16 May 2018, at 19:47, Dani Hak <[log in to unmask]> wrote:
>>>
>>> Hi All, I am starting to use MELODIC tool to analyze resting state networks. I am getting a very strange time-series signature for all the components: very large spikes that appear towards the end of the TS. They don't appear in the data after pre-processing. Has anyone seen the attached a signature? It looks like an artifact of the analysis. Thanks, Dani. <screenshot.png>
>>
>>
>
>
> ---------------------------------------------------------------------------
> Stephen M. Smith, Professor of Biomedical Engineering
> Head of Analysis,  Oxford University FMRIB Centre
>
> FMRIB, JR Hospital, Headington, Oxford  OX3 9DU, UK
> +44 (0) 1865 222726  (fax 222717)
> [log in to unmask]    http://www.fmrib.ox.ac.uk/~steve
> ---------------------------------------------------------------------------
>
> Stop the cultural destruction of Tibet
>
>
>
>
>
>