Training the TOM network on new data

[aritche]

Hi,

I am trying to generate TOMs from the .trk files provided on Zenodo (link). Having a look at the "Train your own model" instructions in the repo's readme file, it doesn't appear to generate TOMs anywhere throughout the steps.

Is there a file similar to TractSeg/resources/utility_scripts/trk_2_binary.py but for generating TOMs? (e.g. trk_2_TOM.py)

[wasserth]

This is a bit more tricky unfortunately. But the following two issues can tell you how it works in general:
#82
#92

I will add the links to those issues to the "Train your own model" sections.

[aritche]

Thank you very much.

When I generate TOMs via the MitkFiberDirectionExtraction script, they are not as 'smooth' as the TOMs generated by the pre-trained TractSeg network. Is this to be expected? or am I using inappropriate parameters?

The command I am using is: sh MitkFiberDirectionExtraction.sh -i ./CST_left.trk -o ./result_TOM --mask ./CST_left_mask.nii.gz --athresh 10 --peakthresh 0.1 --numdirs 1 --normalization 3 --file_ending .nii.gz

Result from the script (CST_left):

Result from the pre-trained TractSeg network (CST_left):

[wasserth]

That is expected. The TractSeg TOMs are smoother and better than the peak maps from MitkFiberDirectionExtraction.sh