robbert-harms/MDT

MDT kurtosis vs DKE

droediger opened this issue · 2 comments

droediger commented

The same subject was run through HCP (4.0.1) Preprocessing followed by both MDT's kurtosis fitting and DKE kurtosis estimation.

Looking at the axial kurtosis maps (I assume "KurtosisTensor.AK.nii.gz" is axial kurtosis), the MDT output has much higher values across the white matter than the corresponding DKE output. Why are the results this different?

MDT AK:
mdt_ak

DKE AK:
dke_KAX

The mean kurtosis (MK) maps look similar across MDT and DKE but with a more dark voxels (indicating failed model fitting?) in the MDT output. Is this due to some smoothing or robust fitting that is included in DKE but not MDT?

MDT MK:
mdt_mk

DKE MK:
dke_mk

I ran MDT like so:

 mdt-model-fit \
      Kurtosis \
      /input_dir/${sub}/${ses}/T1w/Diffusion/data.nii.gz \
      /input_dir/${sub}/${ses}/T1w/Diffusion/bvecs.prtcl \
      /input_dir/${sub}/${ses}/T1w/Diffusion/nodif_brain_mask.nii.gz \
      --gradient-deviations /input_dir/${sub}/${ses}/T1w/Diffusion/grad_dev.nii.gz
droediger commented

PS looking at the ReturnCodes map, the whole brain is "1" with some speckles of "6" throughout... what am I doing wrong here?

robbert-harms commented

Hi droediger,

These differences are to be expected. MDT uses non-linear optimization methods for all model fitting, similar to [1], except MDT uses a offset-gaussian noise model. Non-linear model fitting may not necessarily be the optimal method for Kurtosis modelling. With a Gaussian noise model the Kurtosis can be estimated using linear optimization methods, which has certain advantages in robustness of fit. With an offset-gaussian or Rician noise model, the standard linear optimization techniques do not work anymore and the next best technique is non-linear model optimization, as done in MDT.

About your other question. A return code of 1 indicates successful optimization, a return code of 6 indicates that it could not readily converge and that the search was terminated at some point in time.

Best wishes,

Robbert

References:
[1]: Constrained Maximum Likelihood Estimation of the Diffusion Kurtosis Tensor Using a Rician Noise Model,
Jelle Veraart, Wim Van Hecke, Jan Sijbers, https://pubmed.ncbi.nlm.nih.gov/21416503/