Question: Default percentile of 95% in HausdorffDistanceImageToImageMetric

[loli]

I wonder if someone could highlight to me the reasoning behind the default percentile of 95% in the Hausdorff distance (HD) ImageToImageMetric.

Intuitively, I would expect the HD to reveal me the largest error e.g. to denote outliers in a image segmnentation task. But with a default percentile value of 95%, the results show something else (I am not quite sure what). Is there any reason behind disregarding the 5% most distant errors?

I personally think a default percentile value of 100% would lead to less surprises and is more consistent with the classical HD definitions (e.g. http://en.wikipedia.org/wiki/Hausdorff_distance).

[prastawa]

Hi,

We use a more robust percentile rather than the max to avoid issues with noisy segmentations. We've had cases where people submit segmentations that look good qualitatively but has a few stray voxels. Using the max penalizes these cases heavily and for that challenge we decided that it wasn't a good approach.

The Hausdorff distance metric should be customized for different applications. Should the clinical task calls for heavier penalty for noisy segmentations, the users can adjust the parameter closer to the 100 percentile.

M

[loli]

Thank you for the answer. I understand your reasoning. Maybe one could add a warning / note somewhere in the source code to make the user aware of the 95% perdentile?

[Marjaneh-T]

Hey Guys,

I was wondering does anyone know how we can change the 95% in percentile Hausdorff distance to the other number?

Thanks