If a template is considered to be of average shape, then this overall
shape may change slightly as more images are added.  The change may not
be all that great, so in practice, it probably does not make so much
difference. I could be completely wrong though.

If you forget about shape changes to the average, then you could
consider just adding to the sufficient statistics used for creating the
average, and re-building the template from these.  DARTEL considers a
model whereby the objective function is between the warped template and
the individual subject.  Because of this, a change of variables is used
to obtain the objective function between the warped individual and the
template.  This causes the Jacobian determinants to enter into the
averaging procedure.  My recent-ish NeuroImage paper says more.

DARTEL of about 100 subjects is reasonably fast though, at least on the
computer that I am now using.  Starting from scratch should require less
than a day of processing.

Best regards,
-John

On Mon, 2009-10-12 at 08:51 +0800, Huiguang He wrote:
> I am also very interest to this question. Even TOM can't create
> Template in the progressive way, I still believe there should be some
> way to do it. As John mentioned if we need some extra matlab code for
> DARTEL. 
> 
> Actually, if template creation is a linear process, it will be very
> easy to solve. However DARTEL is a non-linear transform, may we find a
> sub-optimization method (using linear method to approximate
> non-linear) ?  Since Progressive Template Creation is very useful. For
> example, we have collected about 100 subject in one year, and we want
> to make some statistical analysis, and we create template, VBM,
> DARTEL, ...  After three month, we collect another 30, we want to put
> this data together, if we have progressive template creation method,
> it will save a lot of computing time. Right now, I know some project
> including more than two thousand subjects, so how to effective creat
> template is very useful. 
> 
> What's John's opinion?
> 
> Thanks!
> 
>                   Huiguang 
> 
> Huiguang He, Ph.D.
> Institute of Automation, Chinese Academy of Science 
> 
>  
> 
> 
> On Sat, Oct 10, 2009 at 2:28 PM, Marko Wilke
> <[log in to unmask]> wrote:
>         Hi,
>         
>         
>                 And so with TOM I can't do this progressive generation
>         
>         
>         No.
>         
>         
>                 I must reprocess all
>                 my subjects if I want to add one subject or I must use
>                 other way to create
>                 my template.
>         
>         
>         Yes.
>         
>         However, this is true for every approach you choose towards
>         template creation, though, as not the template but the effect
>         on the processed data is the issue. If you process your data
>         based on different reference data, however defined, then you
>         cannot compare them with each other.
>         
>         Hope this helps,
>         Marko
>         
>         
>         -- 
>         =====================================================================
>         Marko Wilke
>          (Dr.med./M.D.)
>                        [log in to unmask]
>         
>         Universitäts-Kinderklinik              University Children's
>         Hospital
>         Abt. III (Neuropädiatrie)             Dept. III (Pediatric
>         neurology)
>                    Hoppe-Seyler-Str. 1, D - 72076 Tübingen
>         Tel.: (+49) 07071 29-83416                   Fax: (+49) 07071
>         29-5473
>         =====================================================================
>         
> 
-- 
John Ashburner <[log in to unmask]>