#dice-loss

Question 1: Single choice

V-Net segments a prostate that occupies only a small part of each 3D MRI scan. Why did its authors train with a loss based on the Dice coefficient instead of a voxel-wise cross-entropy?

Dice is normalized by the foreground's size, so background voxels cannot swamp it1
The Dice loss is convex in the network weights, so training reaches the global minimum2
The Dice loss can be computed without the ground-truth mask, so unlabelled scans can be used3
Cross-entropy cannot be used when a volume has more than two classes4

Timed test