While working with the SEM dataset, I found that loading sub-rat6_sample-data15_SEM.png results in a 2-channel image. I was concerned about this and found that it has an alpha channel, which should probably be removed.
Also, the rest of the dataset should be examined to avoid this in the future. The training images are expected to have a single channel, so I'm a bit worried of unexpected behavior if there's an additionnel empty channel.