search

AImageLab-zip / ToothFairy

Repository relative to the ToothFairy challenges, MICCAI2023 & MICCAI2024
MIT License
29 stars 7 forks source link

Handling of empty masks in hd95 #5

Closed ykirchhoff closed 2 months ago

ykirchhoff commented 2 months ago

Hi,

the current handling of empty masks in the Hausdorff distance might lead to unwanted behavior. If only the ground truth or the prediction are empty a worst possible value (e.g. the diagonal of the volume) should be assigned and only if both are empty it should be set to 0. Otherwise a false prediction might be assigned the best possible score of 0.

Best, Yannick

LucaLumetti commented 2 months ago

Hi Yannick,

We've been discussing the matter already this morning and agree with your suggestion. The original decision was based on the code from a well-known paper, and I initially trusted their judgment without question. However, I've since reconsidered, and I'll be updating it soon to return the diagonal of the volume.

Best, Luca