doi.org
Accuracy versus Perception: a Benchmark of Deep Learning Models for Single-Image Super-Resolution in Microscopy
Virtual super-resolution deep learning methods provide a powerful solution to overcome the physical and temporal constraints of microscopy imaging. Yet, assessing and choosing an ideal methodological strategy complicates their use in life sciences and creates a lack of trust in these methods. Here we propose an objective comparison of nine popular single-image super-resolution (SISR) models in a collection of publicly available microscopy datasets, including cell components like microtubules, endoplasmic reticulum, and actin, using confocal microscopy, SEM, SIM, SMLM and STED microscopy modalities for fixed and live-cell microscopy data. The proposed models will be assessed quantitatively with a collection of metrics in microscopy and computer vision, and qualitatively by experts in the field. The proposed models will be made accessible through open, user-friendly, containerised notebooks. This systematic assessment of SISR approaches will provide a more comprehensive understanding of these methods' performance and contribute to standardising SISR methods in microscopy.
