Partial Cross Entropy Loss

trained a U-Net model, which

computes loss only on labeled pixels. We simulated sparse annotations by randomly

selecting 100 labeled pixels per image using a function that masks all others as

`IGNORE_INDEX = -1`