Parameter optimisation on high-field MRI

[jron...@gmail.com]

Dear all,

I'm implementing the MEDI-L1 algorithm on an ultra-high-field preclinical MRI probe, and I found the parameter tuning quite challenging and time-consuming as there aren't enough studies reported on similar set-up  that can be used as an reference as we have.

I understand that for the current toolbox (and many QSM algorithms in general), it's suggested to either do a visual examination or the L-curve method for choosing λ, but I'm not entirely sure about the criteria of a 'good' result of visual examination when the images are not standard brain scans. I'm therefore keen to know how to implement the L-curve method for choosing λ, could you please advise me?

Additionally, the smv filtering radius has been affecting a lot of my outcome images, because my MRI scans are generally with smaller FOV (small animal probe) and higher resolution (just below 100 micron). I have a feel of where the radius should be about and its range based on the outcome images - but is there a way to rather objectively analyse/justify the choice of the radius?

Many thanks!

Kind regards,

Jierong

[calb...@gmail.com]

Dear all, 
I am having similar problem in the fine tuning of the parameters. 
I have manually tried a broad range of  λ  values, without success. 
For the smv filtering radius I use values in range of the voxel size. 

@Jierong, Did you manage to solve this problem in the meantime? 

Kind regards, 
Bram 

Op vrijdag 26 februari 2021 om 20:53:52 UTC+1 schreef jron...@gmail.com:

[jron...@gmail.com]

Hi Bram,

I haven't had an answer from the group of Cornell so I revisited their neuroimaging paper on 2012 (doi: 10.1016/j.neuroimage.2011.08.082), so far what I'm doing is to extract the regularisation cost and data fidelity cost from the main MEDI dipole inversion function (MEDI_L1) for each run with a series λ, and plot (supposedly) the L-curve using these numbers. However it means an extended time for parameter tuning (λ), because essentially it's repeating the inversion step with different parameter, i.e. same as running it manually using different λ, and each time takes 15-30 min depending on the dataset. Therefore it takes 2-4hrs to sweep the parameter, and the finer the sweeping step is, the longer it take.

There are other fast reconstruction algorithms that enable less time-consumable/automatic parameter tuning, but the dipole inversion formulations/computations aren't equivalent, so the output (χ) is not necessarily the same.

Best wishes,

Jierong

[calb...@gmail.com]

Dear Jierong, 

I ran the MEDI_L1 code using a wide range of  λ values and plot the L-curve using (x-axis: cost of the data fidelity term (cost_data_vec) ; y-axis: vost of the regularization term (cost_reg_vec)). 
This results in an exponential growth function but I am not reaching the 'L' part of the curve.
I tried  λ  value in between: 10^-5 : 10^4 with different step sizes. 

 

As a result the  χ maps are not looking good. 

could someone please point out if I am doing something wrong. 

Kind regards, 
Bram

Op donderdag 25 maart 2021 om 18:31:00 UTC+1 schreef jron...@gmail.com: