DiceMetric gives weird results

[franciskasara]

Hi!

I am training a MONAI model =SegResNet( out_channels=2) to a segmentation task. I have 2 classes, foreground and background. The groundtruth segmentations/labels are 1 channel images.
I use the MONAI DiceLoss(softmax=True, include_background=False, to_onehot_y=True) which seems to work, it decreases, and the predictions look decent. (include_background is False, because the background is much bigger as the foreground.) However I just can't seem to figure out how to use DiceMetric(include_background=False, reduction="mean", get_not_nans=False).

It either gives an error, or 1 or 0 or numbers above 1.
I read all the tutorials I could find, and try to copy those with no success....

I use the loss like this:
vloss = loss_fn(voutputs, vlabels)

In the tutorials "AsDiscreted" was a common step. I do not want to use MONAI transforms, so this is what I tried:

voutputs_bin=voutputs
voutputs_bin=torch.nn.functional.softmax(voutputs_bin,dim=1)
#voutputs_bin=torch.argmax(voutputs_bin,dim=1)
voutputs_bin = torch.nn.functional.one_hot(voutputs_bin.to(torch.int64), num_classes=-1)
#voutputs_bin=voutputs_bin[:,1:,:,:]
#voutputs_bin=torch.nn.functional.threshold(voutputs_bin,0.5,1)
vlabels_bin=torch.nn.functional.one_hot(vlabels.to(torch.int64), num_classes=-1)

metric(y_pred=voutputs_bin, y=vlabels_bin) 

I included the commented lines, because this, and different combinations of this is what I tried.
After the validation loop I do:

vmetric=metric.aggregate().item()
print(f' dice metric: {vmetric}')
metric.reset()

Could you please tell and explain me, how should I properly use the DiceMetric?

[KumoLiu]

Hi @franciskasara, could you please paste the error message?
And you can refer to the usage here.
https://github.com/Project-MONAI/tutorials/blob/3093dec599f04657a7ae6c7a9c97eba5c910259a/modules/compute_metric.py#L110-L114
Hope it helps, thanks.

[franciskasara]

I have already seen this, and no, it did not help at all. I have tried to use thresholding, like they do in your linked example.

My main concern is NOT the error, it is the metric not working properly. And I'd really like to focus this discussion on how to use it in this case, and the explanation why.

(But when I got the error it was on the aggregating part: RuntimeError: stack expects each tensor to be equal size, but got [1] at entry 0 and [255] at entry 384)

[KumoLiu]

Perhaps posting a small snippet of code that reproduces the issue will make it easier for me to spot the problem. Thanks.

[franciskasara]

Okay, I found the problem.

• In the threshold version the problem was that
  voutputs_bin=torch.nn.functional.threshold(voutputs_bin,threshold=0.5,value=1)
  does not binarize the tensor. It sets everything under 0.5 to 1, and leaves the rest of it untouched.

• In the argmax version the problem was that after
  voutputs_bin=torch.argmax(voutputs_bin,dim=1)
  I did not set the num_classes parameter in DiceMetric to 2, so it messed it up:
  "num_classes – number of input channels (always including the background). When this is None, y_pred.shape[1]"

This is how it works properly: (although it is not going to be 1-DiceLoss.)

voutputs_bin=torch.nn.functional.softmax(voutputs,dim=1)
voutputs_bin=voutputs_bin>=0.5

metric(y_pred=voutputs_bin, y=vlabels.to(torch.int64))

[KumoLiu]

BTW, you can use AsDiscrete in MONAI to convert data into discrete values.

MONAI/monai/transforms/post/array.py

Line 132 in df4cf92

class AsDiscrete(Transform):