2.1.

Two-Color Labeling and Spectral Separation for Photon Budget Optimization

When targeting functional imaging, it is of critical importance that the photons corresponding to the functional signal of interest are directed to the imaging camera with minimal loss, since this signal is usually close to the sensitivity of detectors. Moreover, the variation of the functional signal would be an additional difficulty if it were used to measure aberrations. In our initial proof of concept for structural imaging,²⁰ fluorescence photons were shared equally, using a 50:50 beamsplitter, between the imaging path and the wavefront sensing path for the sake of experimental simplicity. Jumping to functional imaging requires to take action on the optimization of the photon budget for both optical paths.

In this new setup, we demonstrate the combined use of a specific sample preparation and of a dual—or more—color optical setup at the emission of the LSFM to use photons, which are continuously emitted in a specific spectral range for the AO process, whereas photons corresponding to the calcium signal within a second spectral range are directed to the imaging camera with negligible loss. The principle is the following.

While the first step could be seen as a constraint, current genetic methods now perform such sample preparation in routine, for multiple animal models used in neuroimaging. The addition of a stable structural fluorescence signal to the functional imaging also has the benefit of properly locating the functional signals⁴ and has been used as reference to compare the activity level of neurons between different animal samples.²¹ Moreover, structural labeling usually provides continuous fluorescence signal, thus enabling constant feed of photons to the wavefront sensor that is driving the AO process. Details about the adapted setup and corresponding sample preparation are given in the next paragraphs.

2.2.

Optical Setup

The optical setup is based on the previously reported AO-LSFM system, for which full implementation details are described in Ref. 20 and in Fig. S1 in the Supplementary Material, in particular regarding the ESSH wavefront sensor design and metrology performance. Figure 1(a) provides a simplified representation of the setup (conjugation lenses are not represented), focusing on modifications required for efficient functional imaging (see Sec. 2.1). In the excitation path, two lasers (Cobolt 06-MLD-488 nm and Cobolt Jive-561 nm) are coupled inside a monomode fiber, providing adequate excitation wavelengths for the green fluorescent protein (GFP) and its derived variant for calcium imaging GCaMP7b,²² and the red fluorescent protein ChRFP, (see Sec. 2.3, Sample Preparation). A $5\mu \mathrm{m}$ thick light-sheet is formed by scanning the beam with a galvanometric mirror (GVS002, Thorlabs) and sent to the sample through an air objective ($10\times$, numerical aperture (NA) = 0.3, Olympus) and a custom sample holder with No. 1.5 coverslips on its sides. A variable diameter aperture diaphragm placed in the back focal plane of the illumination objective allows to adjust the thickness of the light-sheet. Since the groups of neurons of interest in the present study, involved in the circadian cycle of the adult Drosophila, do not cover the whole imaging FOV of the imaging camera (see Sec. 4, Results), we adjusted the aperture diaphragm to decrease the light-sheet thickness down to $5\mu \mathrm{m}$ when compared to our previous publication ($7\mu \mathrm{m}$). This allowed to improve optical sectioning and SBR of achieved images. Emitted fluorescence is collected through a high-NA water immersion objective ($25\times$, NA = 0.95, Leica). Fluorescence emitted by GFP (for test experiments) or GCaMP (for functional imaging) is shown in green in Fig. 1(a) and is sent to an sCMOS camera (ORCA Flash V2 - Hamamatsu), whereas fluorescence corresponding to structural cherry red fluorescent protein (ChRFP) labeling (in red in Fig. 1(a) is sent to the ESSH wavefront sensor. The ESSH analyzer (from Ref. 20, Imagine Optic) has a custom design and is made of a $17\times 23$ microlens array, with a 5.1 mm focal length, each microlens defining a $43\times 43\mathrm{pixels}$ area on a CMOS camera with a $6.9\mu \mathrm{m}$ pixel size. Fluorescence signals are spectrally separated using a combination of a dichroic filter (Semrock FF556-SDi01) and of bandpass filters (Chroma ET525/50, Semrock FF01-618/50). Before reaching the spectral separation unit, fluorescence signal is reflected by a high stroke, high linearity deformable mirror (DM, Mirao 52e - Imagine Optic) for aberration correction. The FOV is $350\times 350\mu {\mathrm{m}}^2$ on the camera and each microlens form a $130\times 130\mu {\mathrm{m}}^2$ FOV image on the ESSH sensor. Before performing imaging experiments, the static system aberrations were calibrated by measuring the aberrations from a point source, coming from a sample made of sparse $2\mu \mathrm{m}$ beads in a 2% agarose gel at the sample location, using the same direct wavefront sensor with a centroïd algorithm. Then a closed-loop was applied to determine the DM shape that compensates for these aberrations. Finally, an optimization was achieved on the image of a bead to take into account the non-common path aberrations coming from the spectral separation unit and visualize a diffraction-limited Airy disc coming from the bead (see Fig. S2 in the Supplementary Material for additional information). For this final shape of the DM, the corresponding wavefront is measured and recorded to be used as a reference for the ESSH in further AO experiments. The DM shape is recorded as the instrument-corrected shape, to be applied before any imaging. The aberration correction of the instrument requires $<10\%$ of the dynamic range of the mirror and thus allows to correct the aberrations induced by the sample in the linear operating range of the mirror). All results presented in this paper without AO are instrument-corrected so the reported gain brought by AO only comes from the correction of sample induced aberrations. The wavefront calculation procedure has been described in detail in a previous article,²³ and the slopes calculation is based on a Southwell geometry. The slopes calculation process based on intercorrelations follows the pseudo-code described in Anugu et al.²⁴

Fig. 1

(a) AO-LSFM schematic setup for functional imaging, as modified from Ref. 20. Red and green beams are of different sizes for representation purpose only. (b)–(d) Examples of cell bodies of FSB neurons in the Drosophila brain with GFP and ChRFP double labeling. Images acquired on a conventional epifluorescence microscope (Zeiss AxioImager M2). (b) GFP signal used for imaging. (c) ChRFP signal used to perform wavefront sensing and real sub-image associated on the ESSH. (d) Both previous signals merged, the localization is identical. Scale bar: $100\mu \mathrm{m}$.

2.3.

Sample Preparation

Samples consist of freshly dissected adult Drosophila brains with sleep-promoting neurons of the fan-shaped body (FSB) expressing either GFP or GCaMP7b, and ChRFP fluorescent proteins (23E10-GAL4 > UAS-mCD8::GFP, UAS-mCD8.ChRFP or 23E10-GAL4>UAS-GCaMP7b, UAS-mCD8.ChRFP), without any sample processing such as sample fixation or clarification. The emission spectra of the green and red fluorescent proteins are respectively centered at 509 and 610 nm. It corresponds to a spectral separation of 101 nm, which ensures negligible impact of the wavelength dependency of aberrations of our samples²⁵ on the quality of the AO correction. Moreover, the width of the emission spectra of the fluorescent proteins, combined with the position of the emission peak, allows a minimal loss of functional GCaMP7b signal when passing through the chosen dichroic filter.

The following genetic lines of Drosophila flies were obtained from Bloomington Drosophila Stock Center: 23E10-GAL4 (#49032), UAS-mCD8::GFP (#32186), UAS-mCD8.ChRFP (#27391), UAS-GCaMP7b (#80907). Standard genetic crosses were used to obtain flies expressing the dual green and red fluorescent proteins in the same neuronal structures, so the correction of aberrations based on the structural signal is optimized for the functional signal. More precisely, if red and green labeling were not done on the same structures, wavefront measurement based on the red structural signal could lead to a different result than a wavefront measurement based on the green functional signal in the case of large aberrations strongly limiting the isoplanetic patch down to a smaller size than the FOV captured by one microlens. Figure 1 illustrates the achieved sample preparation: it shows the FSB neurons as imaged at 30 to $55\mu \mathrm{m}$ depth using a widefield microscope for [Fig. 1(b)] GFP, [Fig. 1(c)] ChRFP, and [Fig. 1(d)] numerically merged signals, over a $350\times 350\mu {\mathrm{m}}^2$ FOV. The thumbnail on the middle image corresponds to the area of the image captured by each microlens of the ESSH to perform wavefront sensing on an extended scene (see Ref. 20 for details of the wavefront sensing process). FSB neurons project axons into a denser area at about 60 to $80\mu \mathrm{m}$ depth; we also report AO-enhanced images of such neurons later in the manuscript (see Sec. 3.1).

2.4.

Number of Corrected Modes

The choice of the number of aberrations modes to be corrected in the AO process is critical to achieve the best quality of aberration correction while still ensuring the capability of the AO loop to properly converge. Spherical aberration is one major aberration to be compensated for in biological samples, in particular due to the index mismatch with its surrounding medium in microscopy observation, but its sole correction does not ensure optimal image quality.^26,27 Correcting higher order modes provides better performance. However, taking into account high-order modes, corresponding to high-order DM eigenmodes of the interaction matrix from the AO calibration process, induces higher sensitivity to the wavefront (WF) measurement noise, thus higher probability of a divergence of the AO loop.²⁸

Even though such a trade-off regarding the number of DM eigenmodes to be taken into account in the AO process has been widely reported as crucial, there has been poor experimental demonstration of its impact on AO-enhanced microscopy in biological samples. We thus experimentally determined such a trade-off in our setup. With the chosen DM, 52 eigenmodes are accessible for the correction, already corresponding to very high WF spatial frequencies, such as defined, e.g., by Zernike polynomials. We compared on two Drosophila samples, prepared according to the protocol described in Section 2.3, images of neuronal projections exhibiting high spatial frequencies without AO and with AO with different numbers of DM eigenmodes. When keeping all 52 modes, the AO loop starts diverging after $\sim 10$ iterations for a high gain value of the AO feedback loop (0.8). Figure 2 presents a comparison between no AO, AO using 16 and 36 DM eigenmodes, on the same FOV, 60 to $80\mu \mathrm{m}$ deep, for two different areas of neuronal projections. For each case, a representation of measured aberrations is given by the values of the first 15 Zernike coefficients, up to the fifth-order spherical aberration. The Zernike coefficients beyond mode 15 are negligible and are therefore not shown here. Without AO, measured aberrations [Fig. 2(a)] confirm results from previous publications,^19,26 i.e., third-order spherical aberration (eighth Zernike coefficient) being dominant but accompanied by mostly third-order aberrations. AO correction based on 16 DM eigenmodes [Fig. 2(b)] show strongly reduced Zernike values but only provides a moderate gain regarding image quality. Interestingly, fifth-order spherical aberration (15th Zernike coefficient) shows in this case an increased value after AO (confirmed by the corresponding wavefront map shown in Fig. S3 in the Supplementary Material), that is likely to explain the limited visual gain on images. We attribute this behavior to the fact that the DM eigenmodes do not correspond to Zernike polynomials. DM eigenmodes are indeed linked to the membrane shape and the spatial distribution of actuators, and even in the case of the Mirao 52e DM exhibiting a circular membrane with regularly distributed actuators underneath, there is no single DM eigenmode perfectly describing the fifth-order spherical aberration, this discrepancy increasing with the order of the aberration to be corrected. For the DM to correctly describe the fifth-order spherical aberration, a combination of several DM eigenmodes, including high orders, is required. As spherical aberration of various orders is very present in depth, the restriction of the number of DM eigenmodes to 16 strongly limits the capability of the AO loop to compensate the fifth-order spherical aberration. This hypothesis is confirmed by the third experiment [Fig. 2(c)], showing the result of AO correction using 36 DM eigenmodes. In this case, there is a visual increase of image quality, regarding image contrast, and all first 15 Zernike coefficients including the fifth-order spherical aberration are reduced to values close to zero. This improvement is confirmed quantitatively by line profiles presented in Fig. S3 in the Supplementary Material. For the two experiments involving AO, convergence of the loop was ensured after five to six iterations, and stability was properly maintained afterwards.

Fig. 2

Images of neuronal projections of FSB neurons between 60 and $80\mu \mathrm{m}$ deep in two different freshly dissected Drosophila brains (a) without AO, and with AO with (b) 16 modes or (c) 36 modes. Scale bar: $20\mu \mathrm{m}$. Graphs represent absolute value of Zernike modes from 4 (astigmatism 0) to 15 (fifth-order spherical aberration) corresponding to images. Associated wavefront and full FOV for top image is presented in Fig. S3 in the Supplementary Material.

3.1.

Ex-Vivo Quantification of AO Performance

We first demonstrate the performance of our new AO-LSFM setup on ex-vivo, freshly dissected adult Drosophila brain samples, with a structural dual color labeling and prepared according to the previously described procedure. Neuron cell bodies [Fig. 3(a)] and neuronal projections [Fig. 3(b)] were imaged at respective approximate depths of 30 to $55\mu \mathrm{m}$ and 60 to $80\mu \mathrm{m}$. Exposure times were 200 ms on the imaging camera (GFP) and 100 ms on the ESSH wavefront sensor (ChRFP). For each structure of interest in (a) and (b), Fig. 3 shows images before (left) and after (right) AO correction, demonstrating a strong gain in image quality. More specifically, sub-images show the corresponding measured wavefronts, that drop from 256 to 34 nm root mean square (RMS) in Fig. 3(a) and from 377 to 37nm RMS in Fig. 3(b), which in both cases correspond to a correction better than $\lambda /15$ RMS as compared to the Maréchal criterion defining diffraction-limited imaging at $\lambda /14$ RMS residual wavefront error. From the measured wavefronts, it can be observed that, as expected, spherical aberration is dominant—together with astigmatism—before the AO process. To quantify the gain brought by AO, we computed the average radial Fourier transform for each image [top graph in the middle of Figs. 3(a) and 3(b)]: it shows a clear enhancement of medium spatial frequencies up to 50% in Figs. 3(a) and 30% in 3(b). This enhancement is quantitatively confirmed by the profiles corresponding to the white arrows in Figs. 3(a) and 3(b) (bottom graph on the right). Profiles show a clear contrast enhancement by a factor of more than 2 in both cases, and a corresponding increase in terms of spatial resolution. The latter is particularly visible in Fig. 3(a): since GFP corresponds to a membrane labeling, AO provides a better view of the nucleus that corresponds to a dip in the profile (and a dark, round structure in the middle of the cytoplasm in the fluorescence image). In each case, the correction was achieved after a maximum of 6 iterations with a gain of 0.6, which corresponds to a typical correction time of 500 ms, mainly limited by the exposure time of the wavefront sensor.

Fig. 3

Image of freshly dissected Drosophila brains with FSB neurons expressing GFP and ChRFP, and associated measured wavefronts before and after AO correction expressed in $\mu \mathrm{m}$. (a) Cell bodies at an approximate depth of 30 to $55\mu \mathrm{m}$. (b) Full FOV of neuronal projections at an approximate depth of 60 to $80\mu \mathrm{m}$ with a zoom on the corrected area. For (a) and (b), in the right part, the top graph shows the average radial Fourier transform of the images before and after AO correction, and the bottom graph shows a profile corresponding to the line between the white arrows before and after AO correction.

3.2.

AO-LSFM for Enhanced Calcium Neuroimaging in an Ex-Vivo Adult Drosophila Brain

We then show the interest of our direct wavefront sensing AO strategy using ESSH for calcium neuroimaging by demonstrating the gain in both sensitivity and spatial resolution in the images obtained with our AO-LSFM setup, at an approximate depth of $40\mu \mathrm{m}$ in an ex-vivo Drosophila brain. The fluorescence signal in the images in Fig. 4 corresponds to emission of GCaMP7b that is proportionate to the native activity levels of sleep neurons at the time of the experiment.^29,30 The fluorescence intensity levels recorded here are therefore consistent with those encountered in current functional neuroimaging experiments to map neuronal activity and ensure the relevance of this proof of concept. Exposure times of the imaging camera and the wavefront analyzer camera were 100 and 200 ms, respectively. The wavefront is measured at $41\mu \mathrm{m}$ ($\pm 10\mu \mathrm{m}$) in-depth and the correction of the aberrations provides a gain in signal of a factor 2, clearly visible in the images of Fig. 4(a) (top) and confirmed by the profiles on the two bright left neurons shown in Fig. 4(b) (bottom). The signal enhancement can be observed at different depths [Fig. 4(a) (bottom)], over a thickness of $10\mu \mathrm{m}$ (see the volume reconstruction in Video 1). This gain in intensity brought by AO directly corresponds to a gain of sensitivity of the imaging setup. Such a gain is of critical importance for functional experiments, since it provides the capability to detect and monitor neurons with lower activity, i.e., with faint signals close to the background level. An example of such neurons can be visualized in Fig. 4(b) (top), as indicated by the red arrows. Before aberrations correction [Fig. 4(b) (top left)], the intensity level of these neurons is of the same order of magnitude than the surrounding signal of brighter cells, whereas in Fig. 4(b) (top left), two neurons can be clearly identified thanks to the signal increase brought by AO. Zoom of the Fig. 4(b) (middle) confirm the gain in sensitivity where AO allows to reveal low level signals of projections and highlight neuronal connections, indicated with red arrows. It has to be noted that the gain provided by AO is directly linked to the amount of aberrations arising from the sample. The experimental results from Fig. 4 were obtained at an approximate depth of $40\mu \mathrm{m}$. As a consequence, it is expected that imaging deeper would provide a higher gain due to the correction of larger aberrations, that might unfortunately be counterbalanced by scattering at some point, depending on the characteristics of the sample. In our experiments in the Drosophila brain, it was possible to demonstrate a significant gain from AO at depths down to $80\mu \mathrm{m}$ for structural imaging and $40\mu \mathrm{m}$ for calcium imaging. For now, we failed to report significant image improvement using AO at larger depths, which is attributed to scattering starting to be predominant.

Fig. 4

Neurons at $40\mu \mathrm{m}$ ($\pm 10\mu \mathrm{m}$) depth labeled with GCaMP7b without and with adaptive optics. (a) (top) Full FOV image with delimited ROI. Scale bar : $100\mu \mathrm{m}$ (bottom center), zoom on previous ROI at several depths. Scale bar: $20\mu \mathrm{m}$ (bottom left and right), max intensity projection along $x$ ($x_{\mathrm{MIP}}$) of the stack without and with adaptive optics. Scale bar: $5\mu \mathrm{m}$, (b) (top) Zoom on the ROI at $z=41\mu \mathrm{m}$ ($\pm 10\mu \mathrm{m}$) with an adapted colorbar. Scale bar: $10\mu \mathrm{m}$ (center) Zoom on the delimited white rectangle above with a saturated colorbar [lower max value as compared to (top)], resulting in low-level signals being more visible, such as neuronal connections indicated with red arrows (bottom). Profiles without (orange) and with (blue) AO taken between white arrows of top images.