ReScale4DL: Balancing Pixel and Contextual Information for Enhanced Bioimage Segmentation

Abstract

Deep learning has established itself as the state-of-the-art approach for segmentation in bioimage analysis. However, these powerful algorithms present an intriguing paradox regarding image resolution: contrary to intuition, lower-resolution images can yield superior performance for specific image analysis carried out by deep learning. This phenomenon is particularly significant in microscopy imaging, where high-resolution acquisitions come with substantial costs in throughput, data storage requirements, and potential photodamage to specimens. Through systematic experimentation, we evaluate how varying image resolution impacts deep learning performance in cellular image segmentation tasks. We trained popular architectures on datasets downsampled to 6-50% of their original resolution, mimicking acquisitions at lower image magnification, and compared their performance against models trained on native-resolution images. Our results show that segmentation accuracy either improves (by up to 25% of mean Intersection over Union (IoU)) or experiences only minimal degradation (< 5% of mean IoU) when using images downsampled by up to a factor of 4 (25% of the original resolution). This downsampling proportionally increases information throughput while reducing data storage requirements and inference time. With these findings, we contribute systematic guidelines to deep learning practitioners in creating efficient experimental pipelines for image-driven discoveries. This approach improves the sustainability and cost-effectiveness of bioimaging studies by reducing data and computing needs while optimising microscopy techniques.