Fig. 2. ReScale4DL pipeline.

Researchers can optimise the pixel size for a chosen deep learning architecture using their newly acquired and annotated images or on existing datasets, which can be used as a starting point for the optimisation. A series of image rescaling factors is applied to the data, a deep learning model is trained for each rescaled dataset, and different accuracy metrics and features are computed. Once an ideal target metric is chosen, such as the accuracy of the image processing performance, the recovery of an image feature, or a good compromise between the throughput and accuracy, researchers can identify an adequate pixel size and launch their imaging experiments with optimised image resolution. The results presented in this manuscript were obtained by following the ReScale4DL pipeline.