[Castor-users] Fwd: Re: Optimizer parameters and figures of merit

[tmerlin]

Hello Kajal,

The /-opti-stat/ option will output several image-based information for 
each image update (iterations or subsets). Namely, the mean voxel value 
and standard deviation over all voxels in the new image estimation (/New 
image estimate/ line), and the same figures of merit for the difference 
image between the new image estimation and the previous one (/Image 
update step/ line). Just to let you know, you can also configure the 
output number format (scientific/fixed format, and precision ) with the 
option /-onbp./

The first lines display the initial parameters of the /optimizer/ you 
use (MLEM by default). The data space denominator threshold if the 
threshold under which estimated forward values are discarded (it is used 
to avoid division by 0, and could be modified if you have low-stat data, 
leading to very low forward-projeted values). It is used during the data 
updates ( /iOptimizerMLEM::DataSpaceSpecificOperations()/ ). Minimum 
(and Maximum) update factors are some thresholds used for the same 
purposes during the image data step 
(/iOptimizerMLEM::ImageSpaceSpecificOperations() /). All these 
parameters can be modified in the command-line options (using -opti 
MLEM,a,b,c,d   where a,b,c,d are the initial image value, denominator 
threshold, minimum image update and maximum image update values 
respectively), or directly in the MLEM configuration file in 
config/optimizer/.

The /-opti-fom/ option computes some data-space based figures of merit, 
such as log-likelihood and RMSE. As these are computed from the data and 
forward model estimation for each bin, this option is restricted to 
histogram data. There is currently no implementation of image-based 
figures of merit besides global mean and standard deviation 
computations, however it should be quite straightforward to implement 
some in the vOptimizer class (by creating and initializing a variable 
following the example of /mp_imageStatMean/ in vOptimizer.hh and 
vOptimizer.cc files, then implementing its behavior in the 
/vOptimizer::ImageUpdateStep()/ function and outputing the result).

Hope this helps !

Best regards,

Thibaut




-------- Forwarded Message --------
Subject: 	Re: [Castor-users] Problem in crystal index
Date: 	Fri, 25 May 2018 10:00:22 +0200
From: 	Kajal Aggarwal <kajal792 at gmail.com>
To: 	tmerlin <Thibaut.Merlin at univ-brest.fr>



Hello,

I am trying to verify the convergence of reconstructed image for each 
iteration, and to do the same while reconstructing the image I wrote an 
option of "-opti-stat" . (not really clear what this option is doing)

In the log file, because of this option, I got the following lines:

*--> Initial image value: 1
   --> Data space denominator threshold: 1e-10
   --> Minimum image update factor: 0.01
.
.
.
.

   vOptimizer::ImageUpdateStep() -> Start image update
   --> Image update step  | mean: -0.99 | stdv: 0
   --> New image estimate | mean: 0.01 | stdv: 0 | min/max: 0.01/1
.
.
.
.

vOptimizer::ImageUpdateStep() -> Start image update
   --> Image update step  | mean: -0.0099 | stdv: 0
   --> New image estimate | mean: 0.0001 | stdv: 0 | min/max: 0.0001/0.01
*
I tried to look into the vOptimizer.cc file for these parameters, but I 
did not really understand to what factors the 'mean' and 'standard 
deviation' corresponds to. Could you tell me?

I am looking to get the value of likelihood of each iteration, is it 
possible to get the value of likelihood in list-mode?

Thank you.

Regards,
Kajal


-- 
Thanking you.
Yours sincerely,
Kajal Aggarwal
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.castor-project.org/pipermail/castor-users/attachments/20180614/62629198/attachment.html>