Applyin IoU metric for detection thin "bodies"

[andreys42]

Hi, I'm testing your VNet implementation and my case is not a medicine problem : in my train data there are masks that are always looks like thin parts. When estimate model accurcacy my IoU sometimes is really low, because predicted mask located some pixels near of groudtruth , what is basically ok for my case but it gives me really low IoU values (for obvious reasons).
How do you think I can customize my metric to deal with case when predicted (thin) mask located close to true (also thin) mask. Or maybe your know some existed metrics for my case
Thank you in advance