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 samples remains challenging due to optical aberrations and overlapping signals from high-density molecules. Here, we introduce LUNAR, a self-supervised neural-physics framework that overcomes those limitations by jointly optimizing a physical imaging model and a deep neural network. This strategy allows LUNAR to precisely infer 3D molecular positions, photon counts, and aberrations directly from raw, high-density data without prior calibration. Through simulations and experiments, we demonstrate that LUNAR achieves superior robustness and accuracy across diverse imaging conditions, consistently outperforming existing methods. 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 3D SMLM and establishes a general framework for adaptive, data-driven microscopy. Fu and colleagues present LUNAR, a self-supervised neural-physics method that reconstructs 3D molecular positions and optical aberrations directly from raw microscopy data, enabling calibration-free super-resolution imaging in complex biological samples.
Fu et al. (Sat,) studied this question.