Clarification on HD95 implementation

[12k-kushal]

Hi,

I noticed that the current implementation of the 95% Hausdorff Distance (HD95) in medpy differs from the standard formula (wikipedia.org/wiki/Hausdorff_distance). Specifically, medpy computes the 95th percentile of the combined set of directed distances (from prediction to reference and vice versa), whereas the standard definition usually takes the maximum of the 95th percentiles computed separately for each direction.
In my experiments, this is causing a difference of 10mm when compared to the other implementation approach. Could you please clarify if there is a rationale behind this combined percentile implementation? It would be helpful to understand this design choice.

Thank you for your time!

[loli]

Hello, You are indeed correct that MedPy uses a definition of the Symmetric 95% HD that is not shared by everyone. In most cases, I wouldn't expect the difference to be very large between both definitions, but there are edge cases. I think the exact definition should be made clear in the method's docstring and maybe the more usual definition implemented, too. Would you mind supplying a patch? The implementation should be straight forward. Let me know if you have any other questions.

…

On Wed, 2 Jul 2025, 16:06 12k-kushal, ***@***.***> wrote: *12k-kushal* created an issue (loli/medpy#138) <#138> Hi, I noticed that the current implementation of the 95% Hausdorff Distance (HD95) in medpy differs from the standard formula ( wikipedia.org/wiki/Hausdorff_distance). Specifically, medpy computes the 95th percentile of the combined set of directed distances (from prediction to reference and vice versa), whereas the standard definition usually takes the maximum of the 95th percentiles computed separately for each direction. In my experiments, this is causing a difference of 10mm when compared to the other implementation approach. Could you please clarify if there is a rationale behind this combined percentile implementation? It would be helpful to understand this design choice. Thank you for your time! — Reply to this email directly, view it on GitHub <#138>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAGLK7QGAAFIDFFOKAFNQAD3GPRO7AVCNFSM6AAAAACAUCIWASVHI2DSMVQWIX3LMV43ASLTON2WKOZTGE4TMMBUHE4TENI> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>