Draw the tumour.
A compact U-Net trained on the TCGA lower-grade-glioma MRI collection, segmenting the tumour region in a brain slice. Like the emotion model it runs entirely in the browser, and like it, the original repository had no runnable model at all.
Split by patient, not by slice
Consecutive MRI slices of one brain are near-duplicates. Splitting a dataset randomly by slice therefore puts near-copies of the test set into training and buys several points of Dice that do not exist. The 110 patients are split 88/22 and no patient appears on both sides.
Scored on the slices that are empty too
Roughly half these slices contain no tumour. Reporting Dice only on the tumour-bearing ones is flattering, and letting an empty slice score a perfect 1.0 for predicting nothing is worse. Both numbers are published.
Not a medical device
This is a demonstration of a segmentation architecture on a public research dataset. It is not diagnostic, it has never been validated clinically, and it should not inform anybody's care.