The new method, called lattice light sheet activation structured illumination microscopy, or LA-SIM, was described in Nature Photonics by Xue Dong, Quan Meng, Chang Qiao and colleagues working under the supervision of Dong Li across Tsinghua University and the Institute of Biophysics of the Chinese Academy of Sciences. According to the team, the technique improves volumetric resolution by more than a factor of two relative to the diffraction limit, reaching approximately 87 nanometres laterally and 203 nanometres axially. Those numbers matter because the diffraction barrier of visible light, roughly half the wavelength of the light itself, has historically capped what a conventional lens-based microscope can distinguish, blurring anything smaller than about 200 nanometres in the imaging plane.

To understand what LA-SIM does, it helps to unpack its two optical ingredients. The first is lattice light sheet microscopy, a technique introduced by Nobel laureate Eric Betzig’s group in 2014. Instead of focusing light along the same axis used to collect it, a light sheet microscope sweeps an ultrathin sheet of laser light sideways through the specimen, illuminating only the plane currently being imaged. This produces excellent optical sectioning, meaning out-of-focus fluorescence from above and below the plane contributes almost nothing to the image, and it dramatically reduces the total light dose delivered to the sample. That low phototoxicity is what allows living cells and even whole embryos to be imaged continuously without obvious harm.

The second ingredient is structured illumination microscopy, or SIM, developed by Mats Gustafsson around 2000. SIM exploits a clever trick of interference: by illuminating the sample with a finely striped pattern of light rather than uniform illumination, high-frequency spatial information that would normally be lost is converted into lower-frequency moiré fringes that the objective can actually record. Capturing several raw images with the pattern shifted and rotated, then computationally recombining them, doubles the resolution in each dimension. Crucially, SIM is wide-field, acquiring an entire plane at once, which makes it fast enough for live imaging. The catch has always been that SIM’s patterned illumination works best in thin, sparse samples; in thick, densely fluorescent specimens, out-of-focus light contaminates the reconstruction and produces severe artifacts.

LA-SIM fuses the two approaches by using the lattice light sheet not merely to excite fluorescence but to activate it. The system employs reversibly photoswitchable fluorescent proteins, engineered variants of green fluorescent protein that can be toggled between a dark state and a fluorescent state with specific wavelengths of light. In the LA-SIM scheme, a 405-nanometre lattice light sheet first switches molecules on within a thin slab of the specimen. A second, sandwiched illumination step at 488 nanometres thins the activated slab even further along the axial direction. Finally, the detection objective delivers patterned excitation to generate the structured illumination fringes needed for super-resolution reconstruction. Because only molecules within the thin activated volume fluoresce, the out-of-focus background that normally ruins SIM reconstructions in thick samples is largely eliminated before it is ever recorded.

The team also pushed the concept into nonlinear territory. In conventional linear SIM, the resolution gain is capped at a factor of two because the fluorescence response of the dye is proportional to the excitation light. Nonlinear SIM exploits the nonlinear response of photoswitchable proteins: by extending the patterned exposure so that molecules in the bright stripes begin saturating or switching off, the effective fluorescence distribution becomes sharper than the sinusoidal illumination pattern itself, generating higher-order harmonics that translate into further resolution gains. The researchers define a saturation factor as the ratio of the exposure time to the off-switching time of the photoswitchable protein, and their simulations show that this nonlinear variant, combined with axial thinning, delivers the strongest super-resolution performance among all the imaging schemes they compared.

Even with these optical refinements, the raw data posed a problem. Photoswitchable fluorescent proteins, while invaluable for their controllability, tend to be dimmer, less photostable and available in a more limited colour palette than conventional fluorescent labels, and the resulting images suffer from limited signal-to-noise ratios. The team’s answer was a rationalized deep-learning denoising strategy applied in a self-supervised manner, building on earlier work by the same group published in Nature Biotechnology in 2023. Self-supervised approaches can learn to restore images from noisy data alone, without requiring paired clean examples that rarely exist for living specimens. The rationalized framework additionally provides quality control, helping ensure the network enhances rather than hallucinates fine structures, a persistent concern whenever neural networks touch microscopy data.

Those high-quality rationalized deep-learning LA-SIM images then served as training material for something more ambitious: a large-scale three-dimensional transformer model the team calls SRFormer-pMoE. Transformers, the architecture behind modern language models, have been adapted for image restoration in methods such as SwinIR, and the mixture-of-experts design borrowed from models like DeepSeekMoE allows the network to route different kinds of image content to specialized sub-networks. Trained on image pairs spanning five distinct biological structures, including mitochondrial outer membranes, microtubules, F-actin, lysosomes and stress granules, SRFormer-pMoE acts as a general model that can convert comparatively noisy lattice light sheet images into super-resolution reconstructions. In benchmark comparisons against 3D RCAN networks and 2D SwinIR across 1,000 image pairs per structure, the model achieved higher peak signal-to-noise ratios and structural similarity scores. A Bayesian extension of the model also produces confidence maps, flagging regions where the reconstruction should be trusted less, which the researchers used to distinguish fully resolved mitochondrial fission and fusion events from incomplete ones.

The biological demonstrations span scales from single cells to whole embryos. In cultured cells, the system tracked mitochondrial fission and fusion over tens of minutes, followed F-actin cytoskeletal remodelling across entire cell volumes for roughly 37 minutes, and ran a three-colour experiment for 690 time points over about 74 minutes, watching lysosomes, peroxisomes and microtubules interact. Under oxidative stress induced by sodium arsenite, the imaging revealed moving lysosomes mediating the fission of large stress granule condensates, membraneless RNA-protein assemblies implicated in cellular stress responses. In a mouse early embryo labelled for the lysosomal protein LAMP1, the team followed individual lysosomes across the whole embryo for 300 time points spanning two and a half hours. And in the nematode worm Caenorhabditis elegans, two-colour imaging captured the disruption of apical junctions during cell fusion events, as well as the rapid twitching movements of developing larvae.

The practical significance of the depth comparison is hard to overstate. When the researchers imaged mouse embryos at various depths, conventional 3D-SIM using wide-field illumination degraded rapidly with depth, its reconstructions swamped by out-of-focus fluorescence, while the light sheet-based SRFormer-pMoE pipeline maintained image quality far deeper into the specimen. This is precisely the regime where developmental biologists need to work, inside multicellular embryos whose interior cells cannot be flattened onto a coverslip. The team has released the rDL LA-SIM code and the SRFormer-pMoE model on GitHub, and deposited testing datasets and pretrained weights on the Zenodo repository, lowering the barrier for other laboratories to adopt the approach. With patents filed on both the LA-SIM framework and the neural model, and with the field of light sheet super-resolution advancing quickly through parallel efforts such as oblique plane SIM and RESOLFT-based super-sectioning, LA-SIM suggests that the future of watching life at the nanoscale may belong to microscopes and algorithms designed together from the start.

Subject of Research: Lattice light sheet activation structured illumination microscopy for volumetric super-resolution live imaging of cells and embryos

Article Title: Lattice light sheet activation structured illumination super-resolution microscopy

Article References: Dong, X., Meng, Q., Yang, X., Chen, H., Qiao, C., Lin, Y., Zhang, S., Geng, X., Luan, L., Jiang, T., Fu, W., Jiang, A., Xu, W., Guo, J., Wei, R., & Li, D. (2026). Lattice light sheet activation structured illumination super-resolution microscopy. Nature Photonics. https://doi.org/10.1038/s41566-026-02019-6

Image Credits: AI Generated

DOI: 10.1038/s41566-026-02019-6

Keywords: super-resolution microscopy, lattice light sheet microscopy, structured illumination microscopy, deep learning, photoswitchable fluorescent proteins, live-cell imaging, optical sectioning, mitochondria, lysosomes, stress granules, C. elegans, mouse embryo