[Elastix] The Kappa Statistic Metric (and the SSD)
M.Staring at lumc.nl
M.Staring at lumc.nl
Fri Jul 12 09:58:28 CEST 2013
A common mistake with the kappa statistic is to only give it edges, instead of a closed object (filled). Kappa is equal to overlap, and only edges have little overlap usually. Secondly, since these images are usually binary, you need to be careful with the multi-resolution (blurring) and the interpolator.
Alternatively you can perform a distance transform on the segmentation and use SSD, or use SSD directly on the image data. SSD indeed has a range depending on the dynamic range of the input data, which can be easily understood when you take a look at its definition, see eq 2.5 in the manual.
When converting to mhd like this, make sure spacing etc is set correctly. Alternatively, use a proper converter such as pxcastconvert from ITKTools.
From: elastix-bounces at bigr.nl [mailto:elastix-bounces at bigr.nl] On Behalf Of Otilia Cristina Nasui
Sent: donderdag 11 juli 2013 6:38
To: elastix at bigr.nl
Subject: [Elastix] The Kappa Statistic Metric (and the SSD)
I am trying a initial translation on two edge images (micro-MR with fixed
micro-CT) as they are two far apart to run any other transform sensibly.
I could go with the original intensity images and the MI metric...but I thought the edges would run faster and it would give me a chance to try the KS metric. However, each iteration gives me zero for the metric
value. I have tried all possible values for the translational settings.
Could you please let me know how I can begin to address the problem?
PS: I also tried the SSD metric... the values here are in the hundreds? Is there a potential source of error or am I wrong in expecting the metric to have <1 values (as per what I get with the MI metric on the original intensity images.
PPS:I am calling elastix from within Matlab and the images are .mhds (converted from .png converted from raw data sets from Matlab). (These are specimen scans...no DICOM files)
Otilia Cristina Nasui
Elastix mailing list
Elastix at bigr.nl
More information about the Elastix