Mrregister fails

Dear experts,

I’m running the fixel framework and I find an error in the registration step when I use the flag -type rigid_affine_nonlinear.

When I use the version MRtrix 3.0_RC3-15-g9494da8d the following message appears:

mrregister: [ERROR] Linear registration failed, transformation parameters are NaN.
mrtransform: [ERROR] failed to open key/value file "7005/subject2template_warp.mif": No such file or directory
mrtransform: [ERROR] error opening image "7005/subject2template_warp.mif"

and no output is produced.

Using the MRtrix 3.0.0, it appears the following message:

mrregister: [WARNING] final warp computed is not diffeomorphic (negative jacobian determinants detected). Try increasing -nl_disp_smooth or -nl_update_smooth regularisation.

The final output is created but then the warped mask is empty. I suspect is something in the data that makes this step to fail, is the same dataset used here. Any clue about this?

Best regards,


Hi Manuel,

Outright registration failures or negative Jacobians are usually due to issues with the mask (partial or too large) or data problems (intensity normalisation, distortion correction, artefacts, pipeline bugs, …) and unfortunatly require somewhat tedious QC. Developmental data can be challenging to register if images have large differences in brain size but I was able to coregister register equally distributed data ranging from 33 to 44 weeks in a joint template using median aggregation.

Your error message, however, is unusual in that the script should have stopped after mrregister failing and never tried mrtransform. Did you use -continue by any chance?

If you are you able to share your data with me, I am happy to debug the interaction of mrregister and population_template. Also, if you can trigger the issues with a subset or with blurred out data (mrfilter smooth) then there’s, of course, no need to send the originals. With images I could better try to figure out what’s going on with the transformations issue you linked to as well, as verification requires knowing the extent of the masks and header dimensions and transformations as well as the linear transformations.