Figure 2. U-Net is able to mitigate adversarial bias. A: Evaluation of the statistical significance of the induced bias from removing one randomly selected sprout in all images of group 2, resulting in biased group 2b (adversarial approach 1) compared to the U-Net readout trained on this biased data set. B: Assessment of the difference between group 1 and group 2b after adversarial bias was induced by removing two pixels from every sprout of group 2, resulting in group 2b (adversarial approach 2). These results were compared to the readout from the U-Net model trained on this systematically biased data set.