Which loss to implement in training?

[robertderidder]

Hello,
I'm wondering whether I should train on dice loss or if cross-entropy loss also works equally fine for this assignment. Cross-entropy apparently works because the baseline performs okay, but on the coda server the performance is determined using dice loss. Would it be a significant improvement to train on dice loss and is it hard to implement? Hopefully this is okay to ask here

[Nielsvanschaijk]

From my experience using the standard U-Net model from TU/e, training with Dice loss notably improved my results compared to cross-entropy alone, making it worth trying. I experimented with combining Dice and cross-entropy losses (0.5 Dice + 0.5 Cross-Entropy), but this actually reduced performance.

Implementing Dice loss itself is fairly straightforward—it's just a formula that compares overlap between prediction and ground truth, making it intuitive and easy to add to existing code.

[robertderidder]

Thanks that was just what I needed to hear