03 MayDice Similarity Coefficients (DSCs), How Good is “Good Enough”?

If you have a method for automatic segmentation (labeling anatomy) of the human brain in MRI scans, you can test it using a ground truth segmentation by calculating the Dice Similarity Coefficient (DSC).  But what affects DSCs, and how do you know if a DSC value is good or bad?

Some neuroanatomical regions such as ventricles are relatively easy to segment because the borders are well defined and will have relatively higher DCSs.  Other regions, like the amygdala, are difficult and will have lower DSCs.  So there is a range of DSC values that depends on the region of interest.

Neuromorphometrics produces ground truth segmentations by using precisely defined neuroanatomical protocols and our labeled brain scans are created and certified by neuroanatomical experts.  We have labeled 20 subjects from the OASIS repository that were scanned twice so we labeled them twice and then compared the repeat-subject results by registering the two scans and calculating DSCs.  What are “good enough” DSCs?  Here is our current answer.

Post Tagged with , , , , ,

Comments are closed.