Discrepancy of skimage.filters.frangi output between 0.19.3 and 0.22.0 about scikit-image

[lagru]

Thanks for the report. I can reproduce this with Python 3.11 as well. The result changed between 0.19.3 and 0.20.0 from 1.15 to 0.36 (rounded). I haven't checked with git bisect but I am guessing that you are running into changes introduced by #6446. From the release notes

Rewrite skimage.filters.meijering, skimage.filters.sato, skimage.filters.frangi, and skimage.filters.hessian to match the published algorithms more closely. This change is backward incompatible and will lead to different output values compared to the previous implementation. The Hessian matrix calculation is now done more accurately. The filters will now be correctly set to zero whenever one of the Hessian eigenvalues has a sign which is incompatible with a ridge of the desired polarity. The gamma constant of the Frangi filter is now set adaptively based on the maximum Hessian norm (#6446).

In that case, we would consider this a fix and the changed values as intended behavior. Is this a problem for you? Would you have a reason to deem the previous value, 1.15, to be more correct?

[CielAl]

Thanks for the report. I can reproduce this with Python 3.11 as well. The result changed between 0.19.3 and 0.20.0 from 1.15 to 0.36 (rounded). I haven't checked with git bisect but I am guessing that you are running into changes introduced by #6446. From the release notes

Rewrite skimage.filters.meijering, skimage.filters.sato, skimage.filters.frangi, and skimage.filters.hessian to match the published algorithms more closely. This change is backward incompatible and will lead to different output values compared to the previous implementation. The Hessian matrix calculation is now done more accurately. The filters will now be correctly set to zero whenever one of the Hessian eigenvalues has a sign which is incompatible with a ridge of the desired polarity. The gamma constant of the Frangi filter is now set adaptively based on the maximum Hessian norm (#6446).

In that case, we would consider this a fix and the changed values as intended behavior. Is this a problem for you? Would you have a reason to deem the previous value, 1.15, to be more correct?

Hi, appreciate your clarification. My sole purpose to report the case is for certain reproducibility-related tasks and since the envrionment configuration only asks for a range of skimage from 0.19.x to 0.22.x, the skimage was manually installed (with slightly different versions on different machines) and we therefore observe the discrepancy.
I don't have any reason to deem the previous value, but I feel like it might be helpful to other users who encounter the issue in future if we document this with a little more details in the docstring: I did attempt to compare the documentation of frangi between 0.19.x and 0.22 but kinda missed the takeaway of "Earlier versions of this filter were implemented by..." part since it doesn't explicitly address the different computation outputs.

Again thanks for the information!.