Reconstruct Dense Live-Cell Microscopy Images via Learning Continuous Fluorescence Field

Live-cell imaging plays an increasingly important role in developmental biology and drug discovery. The phototoxity of laser light, however, inevitably induces a tradeoff between image quality and cell viability. While learning-based denoising models have been proposed for improving noisy low-quality images, it is practically challenging to obtain high-quality training reference in live-cell imaging. In this work, we present a reference-free framework MicroNeRF for reconstructing dense image stack from sparse slices. By formalizing the fluorescence field as an implicit neural representation, high-quality stack can be posteriorly sampled from continuous manifold. Additionally, we address the interference among fluorescent signals through an adaptive learnable point spread function. Experimental results demonstrate the superiority of MicroNeRF in sparse reconstruction and downstream segmentation task.