Hi, I wonder which sampling strategy you used during training to deal with these highly-unbalanced datasets, especially the melanoma dataset? I didn't find related information in the training script or the paper, so I opened this issue.