Not a medical device. This is a demonstration of a segmentation
architecture on a public research dataset. It has never been clinically validated and
must not inform anybody's care.
01
The slice
Model
Radiologist
Both agree
Loading the model…
02
Slices from held-out patients
Half of these contain a tumour and half do not. The empty ones are the interesting test — a segmenter that hallucinates a tumour in a healthy slice is worse than useless.
03
How it was measured
The split is by patient, not by slice. Consecutive slices of one brain are near-duplicates, so splitting randomly by slice puts near-copies of the test set into training and buys several points of Dice that do not exist. Both numbers are published: Dice across every test slice, and Dice restricted to the slices that actually contain tumour.