Same with normal image registration, my data is pre and post-image and the ground truth. My dataset has been limited to certain organ and anything outside the organ are 0. I found that the UNet even made the prediction on the slice which only has 0. I am not sure why this happens. My model is simple 7 layer UNet, nothing new at all.
My expectation is that UNet should not make any prediction at the 0 area and keep it 0.
