Hi everyone,
I’m having an issue with tcksift2 for one of my weekly ODF datasets and would appreciate some help understanding where the problem comes from. I have a series of weekly ODF files (weeks 32-44 from developing Human Connectome Project), all generated in the same format and with the same dimensions and voxel spacing.
I ran tractography with:
tckgen -act "${FIVE_TT}" -backtrack -seed_gmwmi "${GMWMI}" -select 5000000 "${ODF}" "${TRACKS}" -nthreads "${SLURM_CPUS_PER_TASK}" -force
and then use
tcksift2 "${TRACKS}" "${ODF}" "${SIFT_WEIGHTS}" -act "${FIVE_TT}" -out_mu "${SIFT_MU}" -nthreads "${SLURM_CPUS_PER_TASK}" -force
For example, for week 41, the SIFT2 output is normal:
Proportionality coefficient after streamline mapping is 0.00071866498228817066 52472 fixels have no attributed streamlines; these account for 21.161548799850042% of the initial cost function
15317 of 181258 fixels were tracked, but have been excluded from optimisation due to inadequate reconstruction
Constant A scaling regularisation terms to match data term is 0.0072410649103759545
Iteration CF (data) CF (reg) Streamlines
46 32.907% 1.520% 4934959
However, week 42 fails:
Proportionality coefficient after streamline mapping is nan
678722 fixels have no attributed streamlines; these account for nan% of the initial cost function
No fixels were excluded from optimisation due to poor reconstruction
Constant A scaling regularisation terms to match data term is nan
Iteration CF (data) CF (reg) Streamlines
10 nan% nan% 4966659
The resulting SIFT2 weights are all 1, and the mu coefficient is NaN. I have also tested SIFT2 without ACT, and I still obtain unit weights, so I don’t currently think ACT is the cause. The important point is that the week 41 and week 42 datasets were processed using the same pipeline and have the same image dimensions/voxel spacing. The behaviour changes abruptly at week 42.
I would be very grateful for any suggestions on how to diagnose the source of the NaN proportionality coefficient.
Thanks