search

PTB-MR / mrpro

MR image reconstruction and processing.
https://ptb-mr.github.io/mrpro/
Apache License 2.0
17 stars 2 forks source link

Remove DCF from iterative SENSE algorithm #481

Open ckolbPTB opened 1 day ago

ckolbPTB commented 1 day ago 
          How important is the weighted objective compared to only using a DCF corrected x0?

The one comparison in the Prussmann paper gives a slight advantage for >30 iterations, disadvantage for fewer.

https://arxiv.org/abs/1902.09657 ( Ong, Uecker, Lustig) also argue that the reweighting of the objective function is suboptimal due to the noise reshaping. the maximum likelihood solution would be without the dcf. they suggest proper preconditioning using different algorithms.

Originally posted by @fzimmermann89 in https://github.com/PTB-MR/mrpro/issues/388#issuecomment-2422275281

ckolbPTB commented 1 day ago

related to #473

fzimmermann89 commented 1 day ago

image fessler

fzimmermann89 commented 1 day ago

@Stef-Martin and me tried it out for his data. We had to change the stopping tolerance for cg, otherwise the non-dcf version terminated after fewer iterations and had more artifacts. after fixed 10 iterations, the subjective quality in our ad-hoc tests was the same with and without dcf in the operator. we required ~20 iterations if not starting with a dcf x0 but with zeros.

There also exist really cheap thus fast approximations for the dcf (pipe& menon or even a simple grid approximation) that should be good enough if only used for x0