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

• DOI: 10.1101/2025.02.16.638552
• Publication Date: 2025-02-21T12:05:51Z
• Authors: Shuang Fu; Wei Shi; Yue Fei; Ruixiong Wang; Tianlun Zhang; Yiming Li

Achieving large depth-of-field super-resolution imaging deep inside samples is often hindered by unknown aberrations and overlapping molecular signals in 3D single-molecule localization microscopy. Here, we present LUNAR, a blind method that simultaneously resolves corrects for using neural-physics model. Through self-supervised learning on data without requiring prior knowledge or accurate calibration, LUNAR synergistically optimizes physical model with neural network to estimate key parameters (e.g., positions, photons, aberrations) of molecules. Its hybrid Transformer effectively handles PSFs varying sizes, achieving theoretically maximum precision consecutive blinking events. Extensive simulations experiments demonstrate consistently outperforms current state-of-the-art methods, reducing error up sixfold the presence overlaps, enabling high-fidelity whole-cell reconstruction mitochondria, nucleus, neuronal cytoskeleton at great depths.