Runtime warnings in debiasRician() when sigma_diff = b0 / SNR becomes zero (divide-by-zero / invalid value)

[Kikubernetes]

Hi, thank you for AMICO.

I encountered runtime warnings in amico/preproc.py during doDebiasSignal=True preprocessing, caused by division by zero in the Rician debiasing step.

Context

• ae.set_config('doDebiasSignal', True)
• ae.set_config('DWI-SNR', 30.0) (required when debiasing is enabled)

The warnings came from expressions using sig2 = sigma_diff**2 in denominators, where:

• sigma_diff = b0 / SNR
• and for some voxels, b0 can be zero or extremely small (e.g., boundary / low-signal voxels)

This leads to sig2 == 0 (or near-zero), then warnings such as:

• RuntimeWarning: divide by zero encountered in divide
• RuntimeWarning: invalid value encountered in divide

Minimal patch tested

I tested the following safeguard in debiasRician():

sigma_diff = b0 / SNR
sigma_diff = max(float(sigma_diff), 1e-8)

Result in my test

• The runtime warnings disappeared.

• AMICO preprocessing + NODDI fitting completed successfully.

• Final outputs were unchanged compared with the original doDebiasSignal=True run (voxelwise diff = 0 in-mask for:

  • fit_NDI.nii.gz
  • fit_ODI.nii.gz
  • fit_FWF.nii.gz)

As an additional sanity check, I also compared doDebiasSignal=False vs True outputs and observed clear structured differences in NDI/ODI/FWF maps (not uniform noise), indicating that the debiasing effect itself remains active and meaningful.
NDI maps (same subject, same slice). From left to right: no debiasing, original debiasing, and patched debiasing. The middle and right panels are visually indistinguishable in this example, consistent with the voxelwise comparison (difference = 0 in-mask).

Suggestion

Would you consider adding a small lower bound for sigma_diff (or skipping debiasing for zero/invalid b0 voxels) to avoid numerical instability in low-signal voxels?

If helpful, I can open a PR with the minimal change.