Live-cell nanoscopy stops burning out: tenfold less light, fourfold faster frames, and mitochondria caught budding

Deep-learning post-processing keeps RESOLFT live-cell nanoscopy running at lower light, pushing both recording duration and frame rate upward. Previously, RESOLFT observation time was capped by photoswitching fatigue of reversibly switchable fluorescent proteins, and frame rate was capped by the number of scanning steps. On fixed-cell vimentin filaments, cutting light dose per frame to about one tenth let the supervised HDNsup extend time-lapse recordings to about five times longer; under sparse sampling, Fast-MoNaLISA raised the frame rate fourfold.

The denoising network restores images from low signal-to-noise input while holding nanoscale spatial resolution. Unsupervised methods Noise2Void and SN2N degraded resolution, while supervised methods CARE and UNet-RCAN tended to smooth cellular structures. Measured by full width at half maximum, HDN and HDNsup gave 56 nanometers and 54 nanometers, close to the 58 nanometer high-signal reference, against 120 nanometers for SN2N, 93 for N2V, 73 for CARE and 70 for UNet-RCAN.

Sparse sampling plus temporal context lets point-scanning nanoscopy trade scanning steps for higher frame rate. Earlier approaches to temporal resolution predicted new timepoints between existing recordings or recovered dynamics during multi-shot imaging, without shortening per-frame acquisition. On vimentin filaments, mitochondria and HIV Gag particles, a 50 percent sampling ratio was the most robust single-frame setting, and adding five frames of temporal context made 25 percent sampling reliable.

The accelerated imaging recorded mitochondrial vesicle dynamics that were previously hard to resolve in living cells. Regular MoNaLISA volume rates were too slow to follow vesicle movement, and to distinguish a vesicle from a 2D section of a mitochondrial tip. At 2 hertz in 2D, protrusions lasting 1.5 to 4 seconds and measuring 80 to 200 nanometers were detected, while 1 hertz and 0.5 hertz missed them entirely; 3D volume rates reached 0.5 to 1 hertz.

The methods target live-cell RESOLFT nanoscopy, in particular the parallelized MoNaLISA implementation, reaching about 2 hertz in 2D and 0.5 to 1 hertz in 3D over a roughly 35 by 35 micrometer field of view at about 60 nanometer lateral resolution. The authors released the LISAI software framework and training examples to ease application to new datasets, and note that subsampled restoration is directly applicable to scanning modalities where each reconstructed pixel corresponds to a defined scan position, such as STED or spinning-disk confocal microscopy.

Denoising reliability depends on signal-to-noise ratio: for rsEGFP2-labeled structures, 10 to 20 percent ON-switching was a good compromise, whereas about 5 percent led to inaccurate predictions or loss of faint structures. The Fast-MoNaLISA sampling ratio must be chosen according to sample signal-to-noise ratio and dynamics; 50 percent was most robust in single-frame mode, while temporal context made 25 percent reliable in dynamic acquisitions. The diversity level encoded in HDN's latent space is currently adjusted manually per dataset, and the authors propose treating it as an explicit hyperparameter with automated optimization.