Hello guys. I am trying to measure the similarity measures between a reference binary ground truth and a segmented binary volume. What are signed surface to surface distance, & unsigned surface to surface distance? Are they assd(Average symmetric surface distance) and asd(Average surface distance)? I cannot find any thing which clearly states their difference. @zivy
Not sure about “signed surface to surface distance” as I expect the distance to always be positive. Also not sure about using a non symmetric computation.
If you look at the Segmentation Evaluation Jupyter notebook you’ll see that we used a symmetric computation (
all_surface_distances = seg2ref_distances + ref2seg_distances). Using a non symmetric approach can underestimate the true distances as multiple points from the segmented volume can be mapped to the same point in the reference segmentation and vice versa.
Terminology recommendation: “reference segmentation”, not “ground truth”. The ground truth is only known in simulation, in all other cases it is a “reference segmentation”. By reference segmentation we mean, closest to ground truth using currently available approaches, possibly gets better in the future. Think one expert segmented the structure and later on we have 20 expert segmentations that yield a different/improved reference.
hello, in that notebook,can you please mention how do I read the matrix? Does the -ve thing represent the negative error?
And the second matrix from left to right is HD, ASSD… Right(in row 1)
Please run the next cells in the notebook, they present the data in tabular format and figures with titles.
Volume similarity which is the only, non standard, evaluation metric is defined at the top of the notebook.