Predictions of nuclei and membrane from the phase image (first row) using models trained without augmentations (third row) are inconsistent with the experimental ground truth (second row), especially in the presence of noise (centre column) or at a different magnification (right column). The predictions using the models trained using spatial and intensity augmentations (see text for details) are invariant to noise and equivariant with magnification. The white box in the in-distribution column highlights the rescue of the lost fluorescence label. The white box in the higher-magnification column shows that the model with augmentations correctly predicts one large nucleus whereas the model trained without augmentation predicts two smaller nuclei. Scale bars, 50 m.
