Change image_utils.py to work with nnunetv2 predictions and ground truth labels with multi-class segmentations.

[NataliaAlves13]

The framework accepts .npz predictions with the "probabilities" key as produced by nnunetv2.
The framework accepts ground truth labels with multi-class segmentations and assumes that the lesion segmentations correspond to label=1.

[joeranbosma]

Thanks for the proposed changes, great to make it compatible with nnU-Net v2!

The change results in predictions being loaded from disk twice (first to try with 'softmax' key, then with 'probabilities' key) for v2. Could you update it to:

# read the nnU-Net format
data = np.load(path)
data = data["softmax"] if "softmax" in data else data["probabilities"]
return data.astype("float32")[1]

I'm a bit worried about the final change, lbl[lbl!=1]=0. This would lead to unexpected behaviour when labels > 1 are used.