Fig. 4

Preprocessing pipelines for datasets. In general, we found through development that very little preprocessing needed to be applied to the raw images before inputting into the model, with normalisation harming the segmentation results rather than helping. For the experiment with a change of microscope, two pipelines were developed due to a change in output resolution of the image: one fixing the objective magnification to that of the training data (e.g. scaled to the same pixel size with no cropping); and one fixing the real pixel size with that of the dataset (padding the image to the same size as the training image prior to scaling). We find that the latter far outperforms the former. Scale bars represent 30 μm