Dice Score

[shreyaspadhy]

Dice Score

[shreyaspadhy]

Yes, that's correct! The average DICE score is printed over the batch that's currently being trained over. However, the DICE Score on 'dice.png' is calculated for that particular image.

[krmerchant]

I guess my question is this: A low Dice 'loss" reflects a better forward propagation as a smaller loss indicates you don't have to adjust your parameters during training. However, for Dice "Score", two samples with good overlap (ie. segmentation map closely matches ground truth) should technically be higher(eg. 1 for max similarity, 0 for minimum similarity). Is that a true statement? If that is the case, then the average DICE "loss" is printed out over the training samples where as the Dice "Score" would be in the 'dice.png' file.

[shreyaspadhy]

You got it right! Given two images, a DICE score of 1 means perfect overlap, and a DICE score of 0 means no overlap at all. However, from losses.py, you can see the DICE Loss is calculated as loss = 1 - torch.sum(dice_eso) / dice_eso.size(0), which is 1 minus the average DICE Score across all images in the batch.