[Rtk-users] Rooster

[Cyril Mory]

The bottleneck is the 4D conjugate gradient, and by far. You can also 
try to reduce the number of main loop iterations (10 should already give 
a nice result, but you may need more for your application).
That, plus reducing the size of your reconstructed volume, plus the 
optimizations I'm working on, should give you a reconstruction time of 
less than one hour.

On 06/01/2016 09:39 AM, Guillaume Landry wrote:
> Good morning Cyril,
>
> Thanks for answering, yes I used 1.2.0. I did 30 main iterations, 4 CG
> and 10 for TV (taken from example). Possibly I could reduce the number
> of TV iterations?
>
> Looking forward to your optimizations.
>
> Cheers
> Guillaume
>
> Dr. Guillaume Landry
> Ludwig Maximilians University (LMU) Munich
> Medical Physics
> Am Coulombwall 1
> 85748 Garching
> Tel:+49 (0) 89 289-14077
> Fax:+49 (0) 89 289-14072
>
> On 06/01/2016 09:34 AM, Cyril Mory wrote:
>> Hi Guillaume,
>>
>> Can you also tell how many iterations you have performed (main loop,
>> conjugate gradient, and TV) ?
>> If you are performing 30 iterations of the main loop, with 4 CG
>> iterations, then considering the size of your data I do not think
>> anything is wrong. Reconstructing a smaller volume will give you a
>> large speedup (time is almost linear with number of voxels), so I
>> would recommend that you try again with the smallest possible volume,
>> i.e. the bounding box of your patient.
>>
>> Are you running the release 1.2.0 version ? If so, note that on the
>> master branch of the git repository, I am adding optimizations for the
>> 4D reconstructions. It is not fully functional at the moment, but I'm
>> currently working 100% of my time on it. I will let you know about the
>> next updates.
>>
>> Regards,
>> Cyril
>>
>> On 05/31/2016 07:48 PM, G.Landry at physik.uni-muenchen.de wrote:
>>> Hi Simon,
>>>
>>> Thanks for mailing list suggestion. My initial question was about
>>> typical reconstruction times for rtkfourdrooster. What I tried is
>>> summarized below:
>>>
>>> -Volume size was 410 410 264 and could be easily reduced by a good
>>> margin.
>>>
>>> -The GPU is quadro M4000 with 8GB
>>>
>>> -about 2300 256x256 projections with shifted elekta panel (so called
>>> M20)
>>>
>>> -10 phases
>>>
>>> -for the motion mask at the moment I just used the FOV mask for first
>>> try.
>>>
>>> --gamma_time 0.0001
>>> --gamma_space 0.0001
>>> - spacing 1 1 1
>>> --niter 30
>>> --cgiter 4
>>> --tviter 10
>>>
>>>
>>> Recon time was about 5-6 hours. I saw about 4 Gb used on the card.
>>>
>>> The image looked nice, albeit with the TV "feel/plastic-y".
>>>
>>> Thanks for your feedback
>>> Guillaume
>>>
>>>
>>>
>>>
>>> Quoting Simon Rit <simon.rit at creatis.insa-lyon.fr>:
>>>
>>>> Hi Guillaume,
>>>> I'm adding RTK user list to this conversation, it's better to have
>>>> these conversations on the mailing list IMO. Can you tell us what's
>>>> the volume size and the GPU?
>>>> Cyril is ROOSTER's dev, maybe he could comment on recon times.
>>>> Simon
>>>>
>>>> On Tue, May 31, 2016 at 5:29 PM, <G.Landry at physik.uni-muenchen.de>
>>>> wrote:
>>>>> Hi Simon,
>>>>>
>>>>> I tried the rooster recon. I used the parameters from the example
>>>>> page. The results looks rather nice but it took several hours to
>>>>> run (5-6). Its a big dataset of about 2000 256x256 projections.
>>>>>
>>>>> Guillaume