Physics-driven self-supervised learning for 3D light-field microscopy (www.nature.com)

Researchers introduced SeReNet, a physics-driven, self-supervised reconstruction network for light-field microscopy (LFM) and scanning LFM (sLFM) that delivers near-diffraction-limited 3D reconstructions at millisecond-level speeds without paired ground-truth data. SeReNet exploits the full 4D light-field measurement (x–y spatial and u–v angular) and accurate spatial-angular PSFs through a three-module architecture: a depth-decomposition stage that builds an initial focal stack via image translation/concatenation, a deblurring-and-fusion stage (nine 3D conv layers + interpolation) that restores high-resolution volumes, and a self-supervised module that forward-projects the estimate along angular PSFs and minimizes projection-to-measurement loss. The model is compact (~195k parameters), generalizes far better than supervised networks and handcrafted RL deconvolution, and can be optionally fine-tuned to mitigate the missing-cone axial limitation at a cost to generalization. SeReNet runs up to ~700× faster than iterative tomography. The significance is practical: SeReNet is far more robust to noise, aberration, and sample motion, and scales to massive datasets—enabling day-long, high-speed subcellular 3D imaging across systems (cells, zebrafish, C. elegans, mice). The team processed >300,000 volumes (tens of TB) in five days versus years with prior iterative methods. By embedding physical imaging priors rather than learning unchecked priors, SeReNet reduces hallucination risk and lowers computational barriers, making high-fidelity LFM more accessible for neuroscience, immunology, and in vivo biology.