Neural-physics adaptive reconstruction reveals 3D subcellular nanostructures over a large depth of field

Abstract

Single-molecule localization microscopy (SMLM) enables volumetric nanoscopy by retrieving 3D molecular positions from engineered 2D fluorescence patterns. However, achieving nanoscale resolution over large axial ranges in complex biological samples remains challenging due to optical aberrations and overlapped single molecules. Here, we introduce LUNAR, a blind SMLM framework that can precisely localize high-density molecules even with inaccurate and highly distorted PSFs. LUNAR employs a self-supervised neural-physics learning strategy that jointly optimizes a physical PSF model and a neural network to infer key molecular parameters, including 3D positions, photon counts, and aberrations. Through simulations and experiments, we show that LUNAR achieves superior robustness to aberrations, high-density performance across diverse PSFs over multiple frames. We showcase its capabilities through whole-cell nanoscopy of mitochondria, nuclear pores, and neuronal cytoskeletons at large imaging depths. By uniting deep learning with physical modeling, LUNAR provides a calibration-free solution for aberration-robust volumetric nanoscopy and establishes a general framework for adaptive, data-driven 3D imaging.