Micro S plit : semantic unmixing of fluorescent microscopy data

Abstract

Fluorescence microscopy is constrained by optical limits, fluorophore chemistry and finite photon budgets, imposing trade-offs between imaging speed, resolution and phototoxicity. Here we introduce $\mathrm{Micro}S\mathrm{plit}$ , a deep learning-based computational multiplexing method that enables multiple cellular structures to be imaged simultaneously in a single fluorescent channel and then computationally unmixed. We show that $MicroSplit$ separates up to four superimposed noisy structures into distinct, denoised image channels, enabling faster and more photon-efficient imaging. Built on Variational Splitting Encoder-Decoder networks, $MicroSplit$ models a posterior distribution over solutions, allowing uncertainty-aware predictions and the estimation of spatially resolved prediction errors from posterior variability. We demonstrate robust performance across diverse datasets, noise levels and imaging conditions, and show that $MicroSplit$ improves downstream analysis while reducing photon exposure. All methods, data and trained models are released as open resources, enabling immediate adoption of computational multiplexing in biological imaging.