Differences between AMICO and classic NODDI toolbox

[Bea9619]

Dear all,

I'm performing some checks, trying to explore differences in analysis with NODDI maps obtained from AMICO or NODDI toolbox.
Before fitting NODDI model, I preprocessed my DWI data with FSL's eddy and topup.
These are the settings I put using AMICO:

amico.setup()
path='/home/psicobiologia/ANALISI/noddi/subj/'
output_path='/home/psicobiologia/ANALISI/noddi/subj/
subject_list=os.listdir(path)
for subj in subject_list:
    ae = amico.Evaluation(output_path=(output_path+subj+'/noddi/modulated'))
    ae.set_config('doSaveModulatedMaps', False)
    amico.util.fsl2scheme(path+subj+'/noddi/dwi_noddi.bval', path+subj+'/noddi/dwi_noddi.bvec')
    ae.load_data(path+subj+'/noddi/eddy_unwarped_images.nii', path+subj+'/noddi/dwi_noddi.scheme', mask_filename=path+subj+'/noddi/hifi_nodif_brain_mask.nii', b0_thr=0, replace_bad_voxels=0)
    ae.set_model('NODDI')
    ae.generate_kernels(regenerate=True)
    ae.load_kernels()
    ae.fit()
    ae.save_results()

Then, I fed NODDI maps into TBSS to obtain the skeletonised images for my subjects, and then I ran a model testing differences between patients and HCs in NODDI indexes.
The results are very similar comparing AMICO and NODDI toolbox, except for ODI, which in the two models has different directions:

NODDI toolbox (contrasts testing differences between the two groups; the following are 1-p values):
ODI_tfce_corrp_tstat1 0.8242 (patients>HCs)
ODI_tfce_corrp_tstat2 0.5932 (HCs>patients)

AMICO:
ODI_tfce_corrp_tstat1 0.6880 (patients>HCs)
ODI_tfce_corrp_tstat2 0.8152 (HCs>patients)

Visualizing the two ODI maps, the AMICO one has 1 in all the CSF voxels, while the NODDI toolbox in the same voxels has values like 0.5.

I would ask why there are these differences and, in general, what model is more solid and preferable to use for voxel-wise analyses.

Thank you so much for any help

Beatrice

[daducci]

Dear Beatrice,

ODI in CSF voxels is not meaningful as no axons are expected in CSF, thus its value is undetermined. The original NODDI algorithm and AMICO follow two different strategies to attribute a value for NDI and ODI in there. In any case, as you are focused on the white matter (i.e., TBSS), you should avoid those voxels in your analyses; have you tried excluding those voxels?
In any case, I would suggest you try setting doSaveModulatedMaps to True and use the modulated mapos, instead. Let me know how it goes.
Best,
Ale

[Bea9619]

Dear Ale,

Thank you so much for your reply.
I agree that running analysis for WM outside a mask of WM skeleton would not be advisable.
I only mentioned the CSF voxels to highlight that, in general, voxel values can differ depending on whether the AMICO or the NODDI toolbox is used for model fitting.
In this regard, how would you explain the fact that the t-test results comparing patients and HCs are so different when using AMICO maps versus those from the NODDI toolbox (considering that analyses were run only on voxels part of the WM skeleton)?

Thank you so much again for your availability and your help.

Best,

Beatrice