Smart scanning for low-illumination and fast RESOLFT nanoscopy in vivo

RESOLFT fluorescence nanoscopy can nowadays image details far beyond the diffraction limit. However, signal to noise ratio (SNR) and temporal resolution are still a concern, especially deep inside living cells and organisms. In this work, we developed a non-deterministic scanning approach based on a real-time feedback system which speeds up the acquisition up to 6-fold and decreases the light dose by 70–90% for in vivo imaging. Also, we extended the information content of the images by acquiring the complete temporal evolution of the fluorescence generated by reversible switchable fluorescent proteins. This generates a series of images with different spatial resolution and SNR, from conventional to RESOLFT images, which combined through a multi-image deconvolution algorithm further enhances the effective resolution. We reported nanoscale imaging of organelles up to 35 Hz and actin dynamics during an invasion process at a depth of 20–30 µm inside a living Caenorhabditis elegans worm.

With k1 and k2 being constants. The equation fitted reasonably well with the measured resolution values. The graph was used to estimate the needed OFF pulse duration for the measurement of the Pex-16-rsEGFP2 labeled peroxisomes ( Figure 5). Based on the literature 1 , we decided a resolution of about 100 nm was enough to resolve the individual peroxisomes, and therefore an off-time of 250 µs was chosen.

Structures occupancy and decision-making
In our approach the increase in recording speed is not constant but specimen-adaptive.
Intuitively the less pixels covered by structures in the field of view the faster scan speed will be is again the slowest, both because it has the longest decision time (150 µs), but also because it uses a bigger PSF, and as such less pixels can be skipped compared to the other probes, as is also illustrated in the scheme (Supplementary figure 1c).

Temporal Evolution of the RESOLFT Point-Spread-Function
The point-spread-function (PSF) of a RESOLFT microscope is not only a function of space, but also a function of time. In particular, the region from which the fluorescence is collected changes as a function of the time t, where t = 0 states the beginning of the RESOLFT cycle within the pixel. Here, as example, we describe the temporal evolution of the PSF of the where Isat_ON is defined as the intensity to reduce 50% the population of the OFF state during the ON-switching step, IsatRead1 as the intensity to reduce 50% the population of the ON state during the probe step; Isat_OFF as the intensity to reduce 50% the population of the ON state during the OFF-switching step; IsatRead2 as the intensity to reduce 50% the population of the ON state during the read-out step. Similar Equations can be derived for the other RESOLFT schemes proposed in the main text. The conventional confocal is instead recorded with no OFF switching and with a longer ON switching pulse (50 us) followed by a blue 488 nm pulse (50 us). The longer ON switching pulse is applied to ensure that most of the rsFPs are driven into the ON state and then read-out.
Importantly, no OFF-switching pulse is applied in this configuration. there is a difference in resolution between the two images, confirming the importance of the "two-photon-like" process. The improved spatial resolution of the decision map allows for a more accurate probing since the effective size of the emission spot is even smaller than in conventional confocal microscopy.

Multi-Image Deconvolution for Smart-RESOLFT
To reconstruct the high-SNR final RESOLFT images we modified a Richardson-Lucy-based multi-image deconvolution algorithm already used in many microscopy techniques, such as, 4Pi 3 , stimulated emission depletion (STED ) 4 and image scanning (IS) microscopy. Here, we report its basic idea and the main equations to obtain the final algorithm, whose input are the series of yl images generated by our time-resolved RESOLFT architecture and the associated point-spread-functions hl, and the output is an estimation of the protein distribution in the sample.
We assume that for any pixel i=1,...,n of any image yl with l=1,...,L, the value (yl)i is the realization of an independent Poisson random variable Yi;l with mean value (Hlx)i, i.e., where: (i) x is a discrete mapping of the protein distribution in the sample (regardless in which state they are, i.e., ON or OFF state); (ii) Hl is the discrete notation for the convolution operator Thanks to the independence of each random variable Yi;l, we can write the probability density distribution as where y ={yl}.
Since we assume to know the probability density distribution P Y (y|x) of the data and the unknown object x appears as a set of parameters, the problem of restoring x can be treated as a classical problem of parameters estimation, which is usually solved through the maximum likelihood approach 5 .In our case, this approach consists in introducing the likelihood function For the sake of simplicity, we move to a vector notation In the data presented in the main paper we show that we can perform particle tracking on the individual peroxisomes at the time series shown in (Figure 5a Peak intensity. Considering the above-mentioned values for STED we find that the RESOLFT peak intensity (14-19 kW cm −2 ) is orders of magnitude lower than STED (89-350 MW cm −2 ).
Therefore, RESOLFT has the potential to reduce photo-bleaching by high order photon interaction, which is known to trigger ROS formation, DNA damage and plasma formation.
RESOLFT imaging applied to green emitting rsFPs such as rsEGFP2 and DronpaM159T, as in this work, requires an average power of 1-10 uW applied for 0.