ReScale4DL: balancing pixel and contextual information for enhanced bioimage segmentation

Abstract Deep learning is the state-of-the-art approach for bioimage segmentation. However, it presents a paradox regarding image resolution: counterintuitively, deep learning segmentation performance can improve with lower image resolutions. This phenomenon is particularly significant in microscopy, where high-resolution acquisitions come with substantial costs in throughput, storage requirements and potential photodamage. We systematically evaluate how image resolution impacts segmentation by training popular architectures on datasets downsampled to 6-50% of their original resolution, mimicking lower-magnification acquisitions. Compared with models trained on native-resolution images, segmentation accuracy either improves (by up to 25% of mean Intersection over Union (IoU)) or degrades minimally (< 5% of mean IoU) when using images downsampled by up to fourfold (25% of the original resolution). Downsampling proportionally increases information throughput while reducing storage requirements and inference time. These findings provide practical guidelines for creating efficient, sustainable and cost-effective bioimaging pipelines that reduce data and computing needs while optimising microscopy techniques.