BPS2026 – Self-supervised learning empowers generative adversarial networks for label-free to fluorescence image translation

Label-free microscopy provides non-invasive access to cellular structures by exploiting intrinsic light properties, as in phase-contrast imaging or the M14 component of the Mueller Matrix microscopy. However, these techniques lack the molecular specificity of fluorescence microscopy, which uses labeled markers to visualize cellular components at high resolution, at the expense of introducing phototoxicity and potential perturbation of natural processes. Recent advances in deep learning, particularly generative adversarial networks (GANs), have enabled the translation of label-free images into fluorescence-like representations. Yet, these models are often trained from scratch due to the scarcity of large-scale paired microscopy data sets, limiting their ability to capture complex cellular structures. In this work, we leverage self-supervised pretraining for GAN-based label-free to fluorescence image translation, to learn robust representations directly from microscopy images. Our approach enhances the model’s ability to capture structural and morphological cues of the cellular structure. We also present a new in-house data set of paired phase-contrast and chromatin-labeled fluorescence images and benchmark our method against the Allen Cell data set. Our experiments demonstrate that self-supervised pretraining substantially improves translation fidelity compared to randomly initialized GANs, underscoring its potential as a scalable strategy for virtual staining in microscopy and advancing the non-invasive study of cellular dynamics.