Make networks more robust against localized artifacts

[constantinpape]

Currently both the network used for the nucleus segmentation as well as boundary and foreground prediction are not robust against localized image artifact. See https://github.com/hci-unihd/antibodies-analysis-issues/issues/82#issuecomment-629273517 for some examples.

I think this is because these artifacts alter the image statistics and mess up the normalization. We should use a more robust normalization procedure. Some ideas:

• use median and mad instead of mean and std
• compute the statistics over the whole well and take the median
• compute the statistics on patches and take the median

cc @wolny

[constantinpape]

Ok, I briefly discussed this with @wolny and for now, we think the best solution would be to determine the normalization parameters over the whole plate (excluding manual outliers!) in a first pass. For this, we should use robust estimators, so median and mad instead of mean and std or quantiles instead of min/max.