AO DIVER: Development of a three-dimensional adaptive optics system to advance the depth limits of multiphoton imaging

Deep tissue imaging has the potential to transform biology and medicine by visualizing the interactions between cells and their environments on a tissue-wide scale. One of the most promising imaging modalities for non-invasive deep tissue imaging with submicrometer resolution is multiphoton microscopy (MPM).^1–3 By using a pulsed infrared laser to image deep into samples, MPM causes less phototoxicity and suffers less from image blurring compared to the shorter wavelengths of light used in confocal microscopy.^1–3 

Scattering and absorption of the infrared excitation light used in MPM are significantly smaller in comparison with the shorter visible wavelengths used in confocal microscopy.¹ Thicker sections of tissue can be imaged due to the improved depth penetration of longer wavelengths, allowing for more accurate measurements of morphology and physical properties.³ Even in highly scattering samples, the spatio-temporal focusing requirements of nonlinear signal generation mitigate the blurring effect of scattering.⁴ In most fresh and fixed tissues, image blurring and background noise limit conventional multiphoton microscopes to depths of hundreds of micrometers.^3–5 Optical clearing of tissue improves imaging depths but requires chemical treatment with organic solvents or ionic detergents,⁶ changing the physical and chemical properties of tissues.

In contrast to the conventional epi-detection geometry, the DIVER (Deep Imaging Via Enhanced photon Recovery) microscope⁷ detects fluorescence light through the sample—on the opposite side of the objective—dramatically cutting the intensity of top-surface fluorescence (Fig. 1).⁸ The wide area DIVER detector also improves signal detection efficiency as a function of depth, due to the increasing numerical aperture.⁸ The DIVER detector mitigates the problem of signal acquisition deep in scattering tissue samples.⁹ However, tissue heterogeneity distorts the propagation of excitation light, resulting in image blurring that grows with the depth of the imaging plane (Fig. 1).

To enhance focusing of laser light into scattering samples and extend the imaging range of MPM, we integrated a deformable mirror (DM) in the excitation path of the microscope—AO DIVER. We developed a wavefront optimization method based on the downhill simplex method^10,11 to simultaneously optimize all n actuators in the deformable mirror in parallel, leveraging the nonlinear dependence of multiphoton excitation fluorescence on intensity as a measure of the wavefront error.^12–14 In contrast to other indirect wavefront sensing approaches,^15,16 our optimization based approach makes no assumptions about orthogonality and allows the deformable mirror to explore a wider range of spatial frequencies. Rather than optimizing the mode amplitude, which constrains the mirror to the initial search set, we initialized the algorithm with a set of orthogonal modes and then allowed the algorithm to adapt the vertices of the simplex arbitrarily—see Fig. 2. This effectively removes constraints in the mirror shape. We constrained the translational modes (tip, tilt, defocus) of the mirror using a modal filtering technique after the algorithm has converged, also illustrated in Fig. 2(d).

Our single-point adaptive optics (AO) method is ideal for imaging deep in highly scattering samples. Other approaches such as image-based optimization^17,18 and direct wavefront sensing¹⁹ require relatively high quality images to begin with. Direct wavefront sensing requires a spatially coherent, bright guide star, but in scattering samples, the guide star signal is deteriorated by the high-frequency phase structure of the sample (Ref. 20 and supplemental figures in Ref. 21).

Using a single point correction with the downhill simplex algorithm at each region of interest is a fast (∼15 s to 65 s) and highly efficient method to improve imaging resolution in depth and minimize losses to stray light. The DIVER detector recovers photons that would otherwise be lost to scattering in the sample, and the transmission geometry minimizes the influence of out-of-focus fluorescence light. In turn, the improved detection efficiency and specificity give high fidelity feedback to the AO optimization algorithm, which further improves the resolution and brightness of the image.

To expand the technique to enable three-dimensional imaging on millimeter scales with a single deformable mirror, we developed a user interface to generate a network of endogenous guide stars. The guide stars are used to create single point wavefront corrections that are interpolated throughout the sample, correcting wavefront aberrations throughout the volume at high speeds.

Details of the DIVER detector construction can be found in Ref. 7. The detector is a photomultiplier tube (PMT) (Hamamatsu R7600P-300) with an 18 × 18 mm² photocathode area. The PMT is optically coupled to a filter/shutter and filled using an index matching liquid (propylene glycol). Two BG39 filters (Schott) of 25 mm diameter are used as windows to seal the chamber and block the excitation laser light.

We use a deformable mirror with 37 hexagonal segments (Boston Micromachines Corp.). Each rigid mirror segment has three actuators that allow for full tip–tilt–piston (X tilt, Y tilt, Z-depth) control of each segment. In total, the imaging system has 3 × 37 = 111 degrees of freedom. To effectively map out the energy landscape, the simplex algorithm is initialized with n + 1 mirror shapes that approximate the Zernike polynomials. Zernike polynomials are often used to constrain the search space in indirect wavefront sensing methods.²² 

The feedback loop is shown in Fig. 3. A femtosecond (fs) laser illuminates the deformable mirror (DM). The DM produces a phase shift, ϕ_DM, which is optically conjugated using a 4F lens system to the X–Y scanners. The scanners are optically conjugated to the back focal plane of the microscope objective. Upon propagation through the sample, heterogeneities in the refractive index throughout the sample cause phase aberrations, which decrease the excitation intensity. Fluorescence light passes from the focal point through a series of absorptive bandpass filters. The signal from the wide-area detector is then digitized and averaged. The algorithm generates the next DM setting, and optimization continues until the fluorescence intensity stops increasing.

The segmented DM has a frame rate of about 40 kHz (25 µs), so that the rate limiting step in the optimization loop is the fluorescence light collection. In the absence of other noise, the confidence that a DM setting ϕ_i improves light focusing scales with the square root of the product of the irradiance value λ_i and the time over which the photons are detected T,

The best strategy for 2PEF optimization in AO DIVER is to keep the laser power low to minimize photobleaching but high enough to maintain an average photon count of at least 0.1 photons/µs. With an acquisition time of 15 ms per setting ϕ_i, this gives a signal to noise ratio of about 38. For small perturbations to the wavefront, it was observed that the optimization routine converges after about 800 iterations. The routine is run three times with added noise each time, to ensure that the algorithm is not stuck in a local minimum. For a typical optimization sequence of 2400 iterations, there are about 3100 total function evaluations on average, which for dim samples takes ≃45 s. For very bright samples, the irradiance can be much higher, allowing a shorter acquisition time of about 5 ms, giving an optimization time of ≃15 s.

To image large 3D samples with AO, we developed a user interface in SimFCS (www.lfd.uci.edu) to help guide the correction process. The focal point is first centered on a local bright spot by the user clicking on the image, which parks the scanner at the desired location. We recorded the absolute x, y, z position relative to the center of the top of the imaging volume and then optimized to find the best wavefront correction that maximizes the brightness and minimizes the excitation volume. Then, the stage is moved to the next z-position, and the next wavefront correction is obtained. If this is done with fine enough z-spacing, we can interpolate the wavefront for arbitrary points between single point corrections, as demonstrated in Fig. 8. The maximum z-step size between corrections depends on the scatterer size, distribution, as well as the numerical aperture of the objective, and the depth of the imaging plane. Using the flat field of beads, we characterized shifts in the field of view caused by optimization in the vicinity of brighter local objects. The possibility to sense brighter objects and shift the field of view depends on the effective focal length of the objective being used (see Fig. 4), but typically, the buffer region from a brighter object should be greater than 1 µm–2 µm when using a 40× 0.8NA water objective and about 3 µm when using a 20× 0.4NA air objective. In other words, if there is a brighter object within 1–2 axial point-spread function (PSF) lengths, the AO system can sense and move toward the brighter region.

The sub-diffraction size of the beads allows us to fit the beads to 3D Gaussian–Lorentzian, a model of the two photon excitation point spread function,²³ 

with

The volumetric data are first binarized by 3 × 3 × 3 median filtering and intensity thresholding. Bead centroids and bounding box are found using Matlab’s bwconncomp and regionprops3 functions. The results from the regionprops3 function serve as the initial guesses for volume fitting, and a least-squares minimization is then performed (using Matlab’s fmincon) to find the optimal parameters I₀, w₀, and w_z to fit the original unprocessed data. The exponential fits used in Figs. 8(d), 8(g), and 8(i) give an estimate of the wavefront aberration. For small deformations in the wavefront, the Maréchal approximation^24,25 gives the result that the intensity of an aberrated focus relative to the intensity of an unaberrated focus is related to the rms phase variation, Δϕ,

1 μm diameter YG fluorescent microspheres (Polysciences, Inc., Warrington, PA) are dispersed into P4 clear silicone using a Vibracell sonicator (Sonics & Materials, Inc., Newtown, CT). After dispersing the microspheres, we added the 9 µm–13 µm glass spheres in various quantities to induce a blurring effect in the sample. Images are acquired with a 20× 0.4NA air objective, 780 nm excitation light, and a 400 nm–600 nm bandpass filter for emission light.

Fresh mouse skin was resected from the back of a K14-GFP (green fluorescent protein) transgenic mouse. The sample was mounted in 1× phosphate buffered saline on a microscope slide and covered with a No. 1.5 cover slip. The sample was imaged within 5 h of resection through the fat side of the skin. The images in Fig. 7 are obtained with a 20× 0.4NA air objective with 900 nm excitation and 575 ± 75 nm bandpass filters for emission light.

A Thy1-YFP HJrs (yellow fluorescent protein-tagged motor neurons) mouse brain, fixed in 10% neutral buffered formalin, was purchased from The Jackson Laboratory (Bar Harbor, ME). The brain was sectioned sagittally with a razor blade to a 2-mm thick slice. The slice was placed between a microscope slide and a No. 1.5 coverslip with agarose to stabilize the tissue and coverslip. After initial attempts to image the ROI shown in Fig. 12, it was determined that residual formalin in the tissue was causing the YFP to photobleach too quickly to optimize the image. The tissue slice was washed for 3 h in a 4 °C solution of phosphate-buffered saline, 5% glycerol, and n-propyl gallate, an antifade compound from Sigma-Aldrich (St. Louis, MO), to decrease the bleaching effect.

The image in Fig. 12 is acquired in a single pass (without tiling) using two-photon excitation of YFP with 950 nm light with a 4× 0.1NA air objective and an emission bandpass filter of 575 ± 75 nm. The wide acceptance angle of DIVER enables a 14 mm wide field of view without vignetting. In Fig. 13, images are acquired using two-photon excitation of YFP with 950 nm light with a 20× 0.4NA air objective and an emission bandpass filter of 575 ± 75 nm. The excitation laser power is increased exponentially as a function of depth to compensate losses to absorption and scattering in the sample.

We developed a simulation of our imaging system to better understand the influence of the DM. The model takes in a user-defined beam profile, typically in the form of a Gaussian beam. We then generated a phase function that defines the aberration of the overall imaging system (sample plus optics). The intensity distribution of the focal volume is then calculated using a Fourier transform of the complex electric field. The maximum value of the focal volume is used to predict the fluorescence signal that serves as the cost function in the optimization algorithm.

AO-enabled image acquisition and optimization software was developed using Delphi. The AO interface was built on top of Globals SimFCS, a program developed in the LFD (Laboratory for Fluorescence Dynamics) for fluorescence microscopy. Data analysis was performed using Matlab and Python.

The imaging depth or the actual focal position (AFP) is calculated using ray optics and estimating the different indices of refraction in the optical setup. The AFP is calculated from the linear encoder readout in the translation stage,

with NA being the numerical aperture, n₁, n₂, n₃ being the refractive indices of the objective immersion fluid, coverslip, and sample mounting fluid, respectively, and c being the thickness of the coverslip in micrometers.

The deformable mirror can produce translational modes, known as tip–tilt and defocus. The total translation depends on the effective focal length of the objective used, the total rotational angle of the deformable mirror, and the radius of curvature of the deformable mirror. The translational modes are illustrated in Fig. 4. To filter out translational modes, the mirror is numerically reconstructed using a lookup table that converts the mirror voltage to deflection. The reconstructed mirror phase, ϕ, is then analyzed by Zernike (⁠$Znm$⁠) transform,

where ρ = [0, 1] and θ = [0, 2π]. Translational modes are $Z1−1$⁠, $Z11$⁠, and $Z20$⁠.

To improve the focusing of excitation light using a high-speed (∼40 kHz) segmented MEMS deformable mirror (Boston Micromachines Hex-111, Cambridge, MA), we implemented a feedback loop (Fig. 3) that uses the intensity of a guide star as a measure of the validity of a wavefront correction.

In the AO DIVER microscope, the guide star can be a fluorescent object of any size that produces two-photon or three-photon excitation fluorescence (2PEF or 3PEF) or has some optical nonlinearity that results in second-harmonic generation (SHG) or third-harmonic generation (THG). In this article, we only show 2PEF results; however, the technique works just as well for 3PEF and has the most robust performance for SHG and THG. The strong performance is due to the photo-stability of coherent nonlinear processes, whereas 2PEF and 3PEF typically experience higher rates of photobleaching, which confounds the optimization process.

The phase and amplitude aberrations introduced by the sample play a significant role in determining the spatial extent and intensity of the point spread function. The characteristics of the point spread function can be described by a composite phase function at the back focal plane ϕ_BF of the objective, made up of system aberrations, ϕ_SYS, and the phase shift from the deformable mirror, ϕ_DM. The sample induced phase shifts, ϕ_Sample, can also be mapped onto ϕ_BF, giving a function ϕ_BF = ϕ_SYS + ϕ_DM + ϕ_Sample. Diffraction-limited imaging is attained when the deformable mirror flattens the phase function ϕ_BF.

In a steady-state system, the 2PEF intensity scales proportionally to the square of the intensity of the excitation light, $Iex(r→,ϕBF)$⁠, and the concentration of the fluorophores, $C(r→)$⁠, as well as a constant α, which accounts for the absorption cross section, quantum efficiency, and detection efficiency,²⁶ 

The calculation of the average fluorescence intensity is simplified by the assumption that the total intensity of the excitation pulse is the laser power P (constant) divided over a volume V, which is changed by the phase at the back focal plane of the system, ϕ_BF. The total number of fluorophores that contribute to the detected fluorescence intensity is determined by the size of the excitation focal volume, V(ϕ_BF), and the concentration of fluorophores inside the volume of observation,

The optimal mirror configuration is determined by the spatial distribution of fluorescent molecules, C(x, y, z), and the focus size and shape, V(ϕ_BF). The dependence of the fluorescence intensity on the concentration and the focal volume means that the algorithm minimizes the volume of illumination and improves optical resolution. However, it can also translate the center of the focus to the location with the highest fluorophore concentration. If this occurs, the deformable mirror can take the shape of the three translational Zernike modes—tip, tilt, and defocus.

One of the challenges with three-dimensional adaptive optics is that each region of the sample has a different wavefront correction if there is any refractive index variation throughout the sample. The AO correction is generally valid for some finite area, known as the isoplanatic patch, over which the wavefront aberration does not change significantly. To produce an AO-corrected z-stack in the most time-efficient manner, we developed a program that allows the microscope user to create a network of AO corrections. To fully cover the volume of interest, the program linearly interpolates wavefront corrections between guide stars. For corrections ϕ_A and ϕ_B at z_A and z_B, respectively, the shape of the mirror ϕ at a location z is given by the linear relation

The interpolation of wavefronts in Eq. (9) and the relationship in Eq. (8) highlight an important feature in implementing the 3D adaptive optics system. If there is a difference in the tip–tilt–defocus modes between z_A and z_B, the deformable mirror continuously shifts the center of the field of view between the two guide stars. The presence of translational modes distorts the images, making objects appear artificially larger or smaller.

Two procedures ensure that the AO system is only optimizing the phase function, and not the spatial location of the focus. The first procedure is shown in Fig. 5. First, the beam is parked on a bright spot in the region of interest [Fig. 5(a)]. Next, to ensure that the beam is parked on a local maximum, we programmed an automated search algorithm that traverses through a small region around the starting point [Δx, Δy, Δz in Fig. 5(b)] to find the brightest point in the region using the XY-scanners and Z-translation stage before starting the AO optimization algorithm. The search algorithm parks the beam on the true center of the guide star, so that the DM does not erroneously translate the field of view during optimization [Fig. 5(c)]. Finally, the calibrated wavefront is saved to a lookup table, which is then used to interpolate the space between guide stars [Fig. 5(d)]. Translation is also constrained during the optimization routine. The program periodically checks the mirror configuration for the presence of translational modes in the mirror setting using the Zernike transform in Eq. (5). If, during the optimization routine, more than 2 µm of translation is detected, it is likely that the guide star is photobleaching or moving, so the routine is terminated. After successful optimization, the field of view is re-centered by the addition of equal and opposite translational modes, if they are detected.

Another challenge in wavefront interpolation is that highly heterogeneous samples require more frequent and varied adaptive optics corrections. For optically clear samples, the wavefront deformation is mainly due to index mismatching between the sample and the immersion medium, which results in spherical aberrations.²⁷ In this case, only top and bottom corrections are needed for good imaging quality throughout the sample. However, for more complex samples with discrete scattering particles (or large, rapid variations in refractive index) throughout the sample, the wavefront deformation varies more quickly and must be corrected at smaller intervals. For linear interpolation to work, the DM setting must also closely follow the true phase correction. For this reason, corrections are done using the top-down approach, starting with a flat wavefront at the top and taking relatively small steps in Z to ensure that the phase perturbation from step to step is small. The optimization is also initialized with the previous step’s correction. In addition, the starting search space is limited to about λ/3 per actuator.

To demonstrate the improvement in both image brightness and resolution as a result of maximizing the fluorescence intensity, we imaged a field of sub-diffraction 500 nm beads spin-coated onto a microscope slide. Setting the deformable mirror to a flat surface (factory setting) resulted in visible astigmatism in the image on the left of Fig. 6(a). After optimizing the intensity of the center bead until convergence, astigmatism is resolved. To quantify the improvement in resolution, images are first binarized to select individual fluorescent beads. We then use the 3D Gaussian–Lorentzian model for a two-photon excitation point spread function to fit the beads and calculate the lateral beam waist as well as the Rayleigh length (also known as the depth of field). There is good agreement between the theoretical model and the actual focal shape [Fig. 6(b)], with beads in the AO off images giving a mean R-squared goodness-of-fit of ∼0.75. The images with AO correction give a near-perfect R-squared value of ∼0.96. On average, the intensity of the beads is increased by a factor of 5. The lateral resolution is improved ∼30% from ∼1.4 µm to a diffraction-limited beam waist of 1.0 µm [Fig. 6(c)]. The axial resolution is also improved from ∼7.0 µm to a diffraction-limited Rayleigh length of 4.4 µm.

The performance of AO DIVER is demonstrated in a skin sample from a transgenic mouse with GFP (green fluorescent protein) tagged keratinocytes (K14). Images are acquired from the fat side of the skin through ∼400 µm of tissue. Skin is a highly scattering sample, with significant variations in the refractive index throughout the tissue, leading to a blurry image [Fig. 7(a)]. At this depth, the isoplanatic patch is relatively small, on the order of ∼20 µm, so lateral wavefront corrections are performed at eight different locations [Fig. 7(d)], using the cell-encoded GFP as guide stars. Each of the deformable mirror settings is saved, and three frames of images are acquired and averaged. A series of eight corrected images are compiled into a single image (AO on) by maximum intensity projection in Fig. 7(b). In Fig. 7(c), the intensity across the inset is plotted, showing that the signal to noise ratio is improved from about 5.8 to over 26.

To quantify the performance of the interpolated volumetric wavefront correction system, samples were prepared with different concentrations of 10 µm diameter glass beads for optical scattering and 1 µm (sub-diffraction) yellow–green fluorescent beads evenly distributed throughout the sample. Wavefront corrections are performed at evenly spaced intervals in depth to produce a network of AO corrections [Fig. 8(a)]. A complete z-stack is acquired with the interpolated wavefronts, as well as with the adaptive optics off measurements, which only uses a system correction—i.e., a flat wavefront in the back focal plane, optimized at the top surface of the sample (Figs. 9–11). In the minimally scattering sample, Fig. 8(b), the Rayleigh length (also known as the depth of field or Z-resolution) stays approximately the same as a function of depth after the AO corrections denoted by the green stars at 50 µm depth. The bead brightness also decreases slowly [Fig. 8(c)]. In Figs. 8(d)–8(f), the scatterer concentration is ∼500 000 glass spheres per ml. Without adaptive optics, the Rayleigh length increases rapidly from a diffraction-limited 4.3 µm to around 8 µm. In contrast, the wavefront corrected beads maintain a Rayleigh length of roughly 4.3 µm throughout the entire volume. Figures 8(g)–8(i) demonstrate the limitations of the AO correction network. The number of AO corrections is higher than that in the lower concentration samples; however, between 400 µm and 800 µm, the correction was performed underneath an air bubble in the sample. As a result, the interpolated wavefronts are only valid for beads close to the guide stars, giving a region with low intensity distorted beads so that Gaussian–Lorentzian fitting did not converge—hence the gap in the data in Fig. 8(i). However, after 800 µm depth, the AO system recovers nearly diffraction-limited resolution. Overall, as long as the spacing between guide stars is correctly chosen, the 3D AO system maintains nearly diffraction-limited resolution throughout the samples, whereas the flat wavefront images rapidly lose resolution due to optical scattering.

A THY1-YFP mouse brain was purchased from The Jackson Laboratory (Bar Harbor, ME) to compare AO DIVER performance to conventional epi-detected imaging performance. Motor neurons in the brain express yellow fluorescent protein (YFP), illuminating the cells throughout the tissue. The image in Fig. 12 is acquired in a single pass (without tiling) with a 4× 0.1NA air objective. The wide area of the DIVER detector (18 × 18 mm²) enables a 15 mm wide field of view without vignetting.

The epi-detected images [Figs. 13(a), 13(e), 13(i), and 13(m)] are compared to the DIVER detected images, both with only a system aberration correction [Figs. 13(b), 13(f), 13(j), and 13(n)] using a 20× 0.4NA air objective. Images from the DIVER detector are also shown with local adaptive optics corrections [Figs. 13(c), 13(g), 13(k), and 13(o)]. The entire volume is then imaged with the complete 3D AO network. Intensity profiles [Figs. 13(d), 13(h), 13(l), and 13(p)] present the fluorescence photon counts across the image in the x-direction, centered over the central neuron in the images. The intensity for each plot is averaged along 15 µm in the y-direction in the image. After a millimeter depth, the epi-detector is unable to resolve details such as the axons or dendrites, whereas the DIVER channel clearly resolves neurons nearly 2 mm into the sample. Signal to noise ratios calculated from Figs. 13(d), 13(h), 13(l), and 13(p) are improved from epi-AO off to DIVER-AO on by a factor of 6, 5.6, 3, and 2, respectively. The decreasing size of the isoplanatic patch as a function of depth is evident when comparing the AO on DIVER images, improving in the brightness of the central neuron, with the radially decreasing image brightness. A side view of the tissue (X–Z maximum intensity projection) shown on the right of Fig. 13 compares the AO off epi-detected image [Fig. 13(q)] to the AO on DIVER detected image [panel (r)] both imaged to a maximum depth of 2.1 mm.

The ability to resolve subcellular structures millimeters deep into tissue will accelerate clinical and scientific discoveries. By noninvasively imaging cells in their native environments with sub-micrometer resolution, we can leverage the broad range of existing light-based biosensors and endogenous fluorophores to understand signaling, morphogenesis, and protein expression and dynamics in complex systems. Improved signal detection and control over light propagation in space and time are paramount to achieve millimeter-scale imaging in biological samples. We implemented AO technology in combination with the DIVER detector to push the limits of deep tissue imaging using skin and brain tissues as example applications.

Our AO method uses a single point intensity approach over image-based wavefront sensing,^16,17 resulting in a speedy optimization that does not require spatial coherence. Regardless of image quality, the algorithm improves the focusing of the excitation light and enhances the image. The system is also relatively insensitive to optical alignment and mirror mode orthogonality, in contrast to conventional indirect and direct wavefront sensing methods.^21,22 The simple wavefront correction algorithm makes adaptive optics more accessible to microscopy labs and users. Three steps are required to improve the optical resolution throughout a volume of interest: point and click on the guide stars throughout the sample and press a button to start the optimization algorithm and then start imaging.

Conventional MPM produces high-resolution images in thinly sectioned samples, but it is often necessary to acquire information throughout tissue at different depths. Wavefront corrections can be extended to three dimensions using wavefront interpolation [Fig. 8(a)]. Interpolation allows for faster wavefront correction, but the density of corrections increases for more heterogeneous samples and with the imaging depth. The correction density depends strongly on the sample, so we used a trial and error approach to improve the image with the lowest number of corrections. AO DIVER resolved the neuron subcellular structure up to ∼2 mm deep into the mouse brain, with an increase in the signal to noise ratio by a factor more than 5 when compared to conventional approaches. Deep imaging, possible with AO DIVER, can be used to study the pathogenesis of neurological disorders such as Parkinson’s disease.²⁸ 

Although laser scanning microscopy is relatively slow, AO DIVER technology can be combined with other imaging modalities, for example, using a single plane illumination technique.²⁹ An extension of our AO system to SPIM will facilitate rapid imaging and could be applied to imaging neuronal activity.

In summary, we have identified and addressed the critical considerations in the development of a 3D AO corrected deep tissue microscope. AO DIVER brings together advanced light shaping technology and efficient signal detection to extend the fundamental limits of deep tissue imaging and make deep tissue imaging accessible and robust. By improving light focusing into the sample and making use of stray excitation light, fluorescence signal generation is more efficient and allows lower laser powers to be used—reducing phototoxicity and sample damage—enabling more physiologically accurate measurements.

E.G., A.D., and S.L. conceived the research; A.D. and S.L. built and aligned the optical setup and microscope; T.G. and S.L. prepared the samples; S.L. developed the AO program; E.G. and S.L. developed the imaging software; S.L. prepared the figures, analyzed the data, and prepared the manuscript; and S.L., T.G., E.G., and A.D. contributed to the final draft.

We thank S. Kuan and M. Plikus for providing the K14-GFP mouse skin sample, T. Vu for rescuing the brain sample, J. Ballesta and Axiom Optics for loaning us an Haso4 wavefront sensor, C. Gohlke for providing the python module lfdfiles, and S. Ranjit, G. Chanan, W.P. Leemans, L. Malacrida, V. Venugopalan, M. Digman, A. Lefebvre, L. Scipioni, F. Palomba, and other LFD members for discussions.

This work was supported under Grant Nos. NIH P41-GM103540 and NIH P50-GM076516. T.G. and S.L. were supported by NSF IGERT BEST Fellowship (Grant No. DGE-1144901).

E.G. and A.D. have a patent on DIVER technology, Apparatus and Method for Light Emission Detection for In-Depth Imaging of Turbid Media, U.S. Pat. 8692998.

The data that support the findings of this study are openly available on GitHub, in repository https://github.com/swleemans/AOdiverData, along with wavefront analysis for Figs. 7 and 13. The program used to run AO DIVER is freely available at https://www.lfd.uci.edu/globals/.