Images without foreground cass in the training

[arthursw]

Is it a problem to have a portion of the training image without any foreground class?
(patients without lesions in the MS lesion segmentation case)

I think it won't help training the network since the dice score and entropy loss cannot be computed for those cases ; is it correct?
Does that mean we should avoid all cases without any foreground class or could it have any benefit?

[FabianIsensee]

Hi, this should not be a problem at all. There might be some stdout warning that you can ignore. I would still include these cases in the training because the model will be more robust if it also sees cases where nothing should be predicted. The dice loss is defined for those cases of course and it will push the network to predict nothing if nothing is present

[arthursw]

All right, thank you very much!

[arthursw]

Just to make sure I understand correctly:

The dice loss is defined as:

        tp, fp, fn, _ = get_tp_fp_fn_tn(x, y, axes, loss_mask, False)

        nominator = 2 * tp + self.smooth
        denominator = 2 * tp + fp + fn + self.smooth

        dc = nominator / (denominator + 1e-8)

So for images (or patches) without foreground case, the dice loss will be 0 no matter if there are false positives or not. That is why I said (too quickly) "the dice loss is not defined for those cases" ; I meant it will not be affected by those cases.

But looking more closely, I understand that batch_dice is set to True (so the dice loss is computed on a batch, not a single patch), meaning if a batch contains a patch with foreground class, and one without, then the dice loss will be affected by the number of false positives. Right?

So if I have too many images without foreground classes, I might end up with batches without foreground class, which would not help the training?

[arthursw]

Ok I was wrong, I didn't see the self.smooth which prevent the loss from being 0 when there is no foreground class! And batch_dice is False in my case. Ok, it's all clear now :)