*Update, Shayan Siddiqui's published research can be found here.
Machines and magic.
After talking with many of you, I'm learning that a core problem with segmentation algorithms is that a lack of annotations renders low significance at the output stage - too much supervision required.
One of the members of this group will soon be publishing medical research that demonstrates the opposite: a DI score of 70%+ using only one or two samples (!) to classify a segment. He said the result was surprising, as if it was magic.
With his permission, I'll share here when he's published.
(Update: Published)