Abstract
The inhomogeneous refractive indices of biological tissues blur and distort single-molecule emission patterns generating image artifacts and decreasing the achievable resolution of single-molecule localization microscopy (SMLM). Conventional sensorless adaptive optics methods rely on iterative mirror changes and image-quality metrics. However, these metrics result in inconsistent metric responses and thus fundamentally limit their efficacy for aberration correction in tissues. To bypass iterative trial-then-evaluate processes, we developed deep learning-driven adaptive optics for SMLM to allow direct inference of wavefront distortion and near real-time compensation. Our trained deep neural network monitors the individual emission patterns from single-molecule experiments, infers their shared wavefront distortion, feeds the estimates through a dynamic filter and drives a deformable mirror to compensate sample-induced aberrations. We demonstrated that our method simultaneously estimates and compensates 28 wavefront deformation shapes and improves the resolution and fidelity of three-dimensional SMLM through >130-µm-thick brain tissue specimens.
Similar content being viewed by others
Main
Fluorescence microscopy is an indispensable tool in visualizing cellular and tissue machinery with molecular specificity; however, in its conventional form, the resolution is limited to 250–700 nm laterally and axially due to the diffraction of light1. Molecular features smaller than this limit cannot be resolved. Super-resolution microscopies such as stimulated emission depletion microscopy2, structured illumination microscopy3 and SMLM4,5,6 have overcome this barrier, allowing biological observations7,8,9,10 well beyond this fundamental limit of light. In particular, SMLM detects individual molecules using photo-switchable or convertible fluorescent dyes or proteins, pinpoints the centers of probes from their emission patterns and reconstructs the molecular centers into a super-resolution image. The unique advantage of SMLM lies in measuring individual molecules without ensemble averaging and, therefore, its potential in molecular counting and ultra-high resolution in both live and fixed specimens11,12. Localization precision as low as 1–10 nm can be achieved in fixed and living cells13,14,15,16.
SMLM in tissues, however, is challenging. One major reason is the distortion and blurring of single-molecule emission patterns (that is, point spread functions (PSFs)) caused by the inhomogeneous refractive indices within the tissue. Such alterations often reduce the information content17 carried by each detected photon, worsening the theoretically achievable localization precision and thus causing resolution loss, which is irreversible by post-processing18. Reversing these sample-induced aberrations requires optical path modifications in a microscopy system, commonly with a deformable mirror or a spatial light modulator, responsive toward each specimen and field of view to adaptively restore the PSFs of single emitters and thus the achievable resolution. This process is known as adaptive optics (AO)19,20,21,22,23.
To guide a deformable mirror to compensate sample-induced aberrations, the distorted wavefront needs to be measured21,22. For point-scanning methods, such as confocal and multiphoton microscopy, the detection focus serves as a ‘guide star’ providing a stable wavefront measurable with a sensor22,23,24,25,26. For wide-field fluorescence modalities, such as structured illumination microscopy, a guide star could be generated by multiphoton excitation27,28, or by embedding a fluorescent bead in the specimen. In contrast, wavefronts from single-molecule emissions, despite their abundance in SMLM experiments, cannot be directly measured as signals from individual molecules blinking stochastically with limited photons, and thus they do not provide the bright and stable signals required for guide stars18. In addition, introducing external guide stars, such as fluorescent beads in SMLM, may drastically increase the fluorescence background, which reduces the detectability of single-molecule emission patterns and generates structured background patterns resulting in localization artifacts18. Besides, if measured directly, wavefronts are composed of not only the aberrated wavefront induced by the specimen, but also the wavefront variations from both in-focus and out-of-focus single emitters at different lateral and axial positions, making it difficult to measure wavefront distortion specific for the SMLM imaging volume.
For this reason, current sensorless AO-SMLM developments29,30,31,32,33 focus on iteratively introducing mirror changes and then evaluating the changes with image-quality metrics. While intensity or sharpness metrics may work robustly for confocal34, two-photon35 structured illumination microscopy36,37,38 and stimulated emission depletion microscopy39, it is difficult to design an image-quality metric that summarizes aberration-related information from a single-molecule blinking frame, while ignoring irrelevant variations, such as intensity, background and molecule positions. In addition to these iterative methods requiring many cycles, including image acquisition and mirror changes, to reach the optimal correction, the optimal metric design varies with structures36. Previous methods for metric-based AO in SMLM provide robust corrections for tissue-induced aberrations only when the target tissue structures are planar or very thin (Extended Data Fig. 1). This is because emission patterns from single molecules at different axial positions result in inconsistent and, in some cases, even opposite metric responses and thus fundamentally limit the efficacy of these approaches for aberration correction in tissues (Supplementary Note 1).
Bypassing the previous iterative trial-then-evaluate processes, we developed deep learning-driven adaptive optics (DL-AO) for SMLM to allow direct inference of wavefront distortion and near real-time compensation. Our trained deep neural network (DNN) monitors the individual emission patterns from single-molecule experiments, infers their shared wavefront distortion, feeds the estimates through a dynamic filter (Kalman) and drives a deformable mirror to compensate sample-induced aberrations. The method, referred to as DL-AO for single-molecule imaging, simultaneously estimates and compensates 28 types of wavefront deformation shapes, restores single-molecule emission patterns approaching the system optimum and improves the precision and fidelity of three-dimensional (3D) SMLM through thick brain tissue over 130 µm, with as few as 3–20 mirror changes.
Results
Design of deep learning-driven adaptive optics
Single-molecule emission patterns generated by individual fluorescence molecules carry information not only about their molecular center positions, but also about the shared wavefront distortion40. The random lateral and axial positions of the blinking fluorescent molecules and their limited photons emitted in SMLM experiments make these emission patterns unsuitable for direct wavefront measurement18. A single-molecule deep neural network (smNet)41 was demonstrated in its capacity to infer wavefront distortions from individual PSFs in simulation and its responsiveness in experimental datasets. Moving from the inference task to active control of a deformable mirror driven by deep learning is, however, nontrivial. Here, we describe our developments in experimental wavefront-based training, stacked estimation networks and stabilized feedback controls through a Kalman filter (Fig. 1) built to allow a robust control and adaptive element correcting 28 aberration modes in near real-time during SMLM imaging, in the presence of complex wavefront distortions, including the distortion induced by refractive index mismatch. Simultaneously compensating a large number of aberration types also enables the capacity of DL-AO in autonomous control of the deformable mirror in response to random and dynamic aberration changes.
Upon detection of SMLM frames, single-molecule-containing subregions are segmented and sent to the network (Supplementary Note 2). Each input subregion goes through a sequence of template matching processes, which are organized as convolutional layers42,43 and residual blocks44 with PReLU activations45 and batch normalizations46 in between, then ‘fully connects’ through 1 × 1 convolutional layers to an output vector of 28 values—amplitude estimates for wavefront shapes in terms of the native mirror deformation modes47 (hereafter referred to as mirror modes). Representing wavefront with coefficients of orthogonal basis helps cut down on the number of outputs and network parameters to be optimized in training. Forming this orthogonal basis directly from native mirror deformations further ensured the coefficients’ accuracy in representing mirror responses. With this consideration, the conversion from mirror modes to Zernike polynomials48—commonly used as the analytical basis to describe aberrations—is dropped to minimize mismatches between mirror responses and Zernike-based wavefront shapes (Supplementary Note 3). The residual differences between theoretical expectations and experimental mirror deformations (Fig. SS3) are incorporated into training data generation.
To build an accurate link between experimentally detected emission patterns and the mirror control with neural networks, it is imperative to train the network with data that match those obtained experimentally. However, experimental training data of single molecules are challenging to obtain, because the ground-truth wavefronts are usually unknown and the extensive variations of the intensity, background and the lateral and axial locations of single emitters, are impractical to cover experimentally. To this end, we simulate wavefront distortions by linearly combining the mirror deformations obtained experimentally in the SMLM system (Supplementary Note 4). We then use the coefficients of these experimental patterns to form the output of the network. The static residue of system aberration after optimizing the microscope system is also incorporated as the baseline of the wavefront shapes. This allows us to efficiently generate millions of training PSFs based on experimentally measured wavefronts with highly accurate training ground truth (Supplementary Note 4, Supplementary Fig. 1 and Extended Data Figs. 2 and 3; 3D-normalized cross-correlation (NCC) value of >0.95, comparing measured PSFs with those generated from network estimation).
Compensating wavefront distortions inferred from PSFs of blinking molecules, we found that the network proposed mirror change fluctuates with non-vanishing uncertainty before/after each mirror update. This uncertainty increases with the network training range, resulting in a trade-off between the compensation range and stability (Fig. SS3). To this end, we drive the deformable mirror by dynamically switching three networks trained with different aberration scales where the transitions between networks are based on the inference uncertainty (Supplementary Note 2). To stabilize network transitions, we used a Kalman filter49 (Supplementary Notes 2 and 5) to reduce the estimation uncertainty by recursively combining wavefront measurements before and after each correction. Due to the uncontrollable availability of single-molecule emission patterns with a high signal-to-background ratio and the evolving PSFs after each correction (Extended Data Figs. 4–6, Supplementary Fig. 2 and Fig. SS3), this process weighs heavily on high-precision measurements against the uncertain ones to ensure stable feedback from the network.
Deep learning-driven adaptive optics characterization
First, we characterized the response accuracy of DL-AO network using controlled wavefront distortions generated by the deformable mirror. These wavefront distortions resulted in aberrated emission patterns, which were then collected and sent to DL-AO network (Methods). By comparing the induced deformation amplitudes with those estimated by DL-AO, we observed that DL-AO network responded toward individual mirror deformations mostly in a one-to-one manner. This behavior was consistently observed with both beads samples and blinking single molecules from immunofluorescence-labeled cell specimens (Extended Data Figs. 2 and 4 and Fig. SS4). At the same time, we also observed that DL-AO sensed changes in other mirror modes besides the one actually being changed, an expected behavior considering that mirror modes are coupled experimentally (Supplementary Note 3). Due to such coupling, mapping between the wavefront shape and mirror mode amplitudes is no longer unique; therefore, we further quantified the network response accuracy through wavefront shape errors and PSF similarities. We observed that independent measurements from DL-AO and phase retrieval18,50 using PSFs of fluorescence beads resulted in nearly identical wavefront shapes with a small difference of 0.13 ± 0.02 rad (mean ± s.d., N = 28) quantified in root-mean-square wavefront error48 (Wrms; Methods and Extended Data Fig. 2). Further, comparing the wavefronts estimated by DL-AO network using single-molecule blinking data (100 PSFs) to those retrieved by phase retrieval from beads, we observed high similarities of 0.83 ± 0.06 (mean ± s.d., N = 28, NCC) and a small wavefront difference of 0.15 ± 0.03 rad (mean ± s.d, N = 28) in Wrms (Extended Data Fig. 4). We observed similar one-to-one responses to mirror changes in both biplane and astigmatism-based setups (Supplementary Note 7), and in an initial investigation on controlling 50 mirror modes simultaneously with DL-AO (Supplementary Note 8). Besides, for the majority of our introduced distortions below 3 radians in Wrms, a single mirror update can already reduce the wavefront error by 50% (Fig. 2e,f and Supplementary Fig. 3). Caused by the nonlinear mirror deformation response to control input51, and the decreased network response amplitudes with the decreasing signal-to-noise level or the increasing network training range (Extended Data Fig. 4 and Fig. SS4), we observed that it usually requires 3–20 mirror updates for full compensation.
DL-AO aims to restore PSFs to the level unmodified by the specimen. To characterize the capacity of DL-AO for PSF restoration, we introduced random wavefront distortions using the deformable mirror and compensated these distortions with DL-AO during SMLM experiments with immunofluorescence-labeled Tom20 in COS-7 cells. Visualizing the raw blinking data during the correction, we found the PSFs became less distorted even after a single compensation, and the mirror shape became stable after ~4 mirror updates (Fig. 2a). Because PSFs from blinking molecules have limited photons and stochastic positions, making them challenging to quantify, we further verified the PSF shape after correction by axially scanning fluorescence beads nearby the compensation areas. Through phase retrieval, we found DL-AO results share a highly similar and flat wavefront shape with the instrument optimum (Methods and Supplementary Note 4), with a residual of 0.29 ± 0.12 rad in Wrms (mean ± s.d., N = 11; Fig. 2b). Comparing the PSFs after DL-AO and the instrument optimum, high similarities of 0.95 ± 0.02 (mean ± s.d., N = 11) were consistently achieved, quantified by 3D NCC (Fig. 2b and Extended Data Fig. 6), and remained 0.96 ± 0.01 (mean ± s.d., N = 11 in NCC) for distortion levels from 0.25 to 2.75 radians in Wrms (Extended Data Fig. 6). Often, this level of restoration was achieved with only 3–6 mirror updates (Extended Data Fig. 6b), and a single mirror update from DL-AO network reduced the wavefront error by 61.2% ± 24.2% (mean ± s.d, N = 11). To drive each mirror update, as few as two subregions containing isolated single emitters were used for DL-AO network estimation, which spent an average of 0.1 s for forward propagation (Supplementary Table 3, Extended Data Fig. 6 and Supplementary Video 4) and made DL-AO suitable for real-time compensation during SMLM acquisition. Besides, simultaneously controlling a large number of mirror modes with high inference speed makes it possible to compensate aberrations in the presence of dynamic changes from the sample structures (Supplementary Video 8). In this direction, as a proof-of-principle demonstration for dynamic aberration correction, we showed that DL-AO can respond to sudden wavefront changes and compensating for randomly induced wavefront distortions, while monitoring single-molecule blinking frames continuously in an autonomous manner (Fig. 2h, Supplementary Videos 6 and 7 and Supplementary Fig. 5).
Next, we evaluated the robustness of DL-AO on compensating different levels of wavefront distortion, from 0.25 to 2.75 radians in Wrms, by assessing the residual wavefront error after correction using both simulation and single-molecule blinking data. After one mirror update, we observed that 51.9% ± 9.3% and 64.3% ± 12.8% (mean ± s.d., N = 165) of the induced level was compensated for experimental and simulated data, respectively (Fig. 2e,f). After 19 mirror updates, the residual level was 0.32 ± 0.02 and 0.08 ± 0.03 (mean ± s.d., N = 165) radians for experimental and simulated data, respectively (Supplementary Fig. 3). This is a substantial improvement, as compared to existing metric-based methods29,30,31,32,33, for example, Robust and Effective Adaptive Optics in Localization Microscopy (REALM)33, which works up to 1 radian at the expense of 10 mirror updates per aberration mode, requiring a total of 330 updates to compensate 11 aberration types (3 rounds)33. In addition, metric-based AO is unstable when imaging volumetric cellular structures (Fig. 2c,d,g, Extended Data Fig. 1, Supplementary Figs. 4 and 5 and Supplementary Video 5). A detailed discussion and quantification of these intrinsic limitations of metric-based methods can be found in Supplementary Note 1. We note that when the PSF is in focus, metric-based AO works robustly to compensate aberrations, and thus metric-based AO was used in this work to perform system flattening in obtaining an instrument optimum pupil function for training DL-AO networks.
Validation through tissue and cell specimens
Inhomogeneous refractive indices within cells and tissues redirect and scatter light. In particular, the mismatches between refractive indices in sample media and objective immersion media reduce the shape modulation of the single-molecule emission patterns axially and broaden the focus laterally (Fig. 2d), increasing the localization uncertainty in all directions and thus worsening the resolution of SMLM. Such resolution deterioration becomes more drastic with an increasing imaging depth18.
Here, we demonstrate the capacity of DL-AO in compensating index mismatch-induced aberrations using constructed specimens from \(\sim\)35 µm to 134 µm in thickness with water-based imaging media. Imaging immunofluorescence-labeled Tom20 in COS-7 cells through such thickness without AO correction, the super-resolution images of Tom20 proteins showed nearly no axial distributions (visualized by color differences; Fig. 3a and Extended Data Figs. 8a and 9a), a consequence of the severe lack of shape modulation along the axial direction due to the large imaging depth. While the raw data for both cases in the comparison were acquired in an interleaved manner without and with AO (Methods), DL-AO reconstruction showed the expected outer membrane contours of mitochondria, and without AO the reconstruction displayed notable artifacts (Fig. 3b,c). Zooming in on the lateral dimension, we observed the aggregations of Tom20 proteins, known to form clusters52, when aberrations were corrected by DL-AO. In comparison, without DL-AO, the lateral reconstruction of Tom20 distribution is diffusive (Fig. 3d,g), as a result of deteriorated lateral resolution through the large imaging depth. These resolution contrasts without and with DL-AO are consistently observed with different samples (Fig. 3e–g and Extended Data Figs. 8 and 9).
Next, we illustrate the mechanism behind such resolution improvement (Fig. 3h–k) by looking at the PSFs and pupil function, which summarizes how the sample together with optical system modulates the collected light, before and after AO. In comparison to the near-uniform distribution of magnitude and phase in the pupil obtained from an in vitro bead, wavefront (phase in the retrieved pupil) showed substantial radial variations and increased phase wrappings at large radial positions (Fig. 3h and Extended Data Figs. 8d and 9d). As a result, the PSFs at different axial positions throughout a 2-µm axial range remained nearly invariant (Fig. 3j). Such loss of PSF shape modulation results in localization artifacts where identical axial positions are falsely assigned to molecules despite their axial distributions. In contrast, DL-AO restored the flatness of the wavefront, resulting in PSFs that are highly similar to the instrument optimum (Fig. 3h,j and Extended Data Figs. 8d and 9d). These improvements in PSF sharpness and modulation explain the resolution improvement after DL-AO (Fig. 3c,d,f,g and Extended Data Figs. 8c and 9c) and were further quantified statistically showing increased Fisher information content per photon upon DL-AO correction (Fig. 3k).
We further demonstrated DL-AO on arbitrary tissue-induced aberrations by imaging through 200-µm-cut unlabeled brain sections resolving membrane of mitochondria using immunofluorescence-labeled Tom20 in COS-7 cells (Fig. 4). Without DL-AO, our observation is consistent with that through water-based cavities where the information of Tom20’s axial distribution is lost even with the in situ PSF model (Fig. 4a). Further deterioration was observed both laterally and axially (Fig. 4a,f) using an in vitro PSF model with theoretical index mismatch aberration incorporated. With DL-AO, the 3D reconstruction showed improved resolution, where such improvement could be visualized laterally by the distinct Tom20 protein clusters and axially by the mitochondria membrane contours (Fig. 4b–e).
Amyloid-β fibrils in 125-µm-cut mouse brain sections
The 3D structures of amyloid-β (Aβ) fibrils are a focus of interest in the studies of Alzheimer’s disease and are of particular importance with the success of amyloid-directed therapeutics53,54. Visualizing the formation and aggregation of these fibrils within the brain has been limited by the notable resolution loss when imaging through tissues. With DL-AO adaptively optimizing single-molecule emission patterns during SMLM imaging, we can now clearly resolve the organization of immunofluorescence-labeled Aβ fibrils in 125-µm-cut brain sections from 5XFAD mice, a transgenic Alzheimer’s disease model that exhibits robust amyloid plaque pathology similar to that found in the human Alzheimer’s disease brain55 (Fig. 5). We imaged Aβ fibrils through these thick brain tissues without and with DL-AO in an interleaved manner. We observed improved resolution in both axial and lateral directions with DL-AO compared with fibrils imaged without AO (Fig. 5b). Importantly, driven by DL-AO, SMLM reconstruction revealed the 3D organization of individual amyloid fibrils entangling and forming the plaque. However, while without DL-AO, the resolution deteriorates, making the intricate fibril ultrastructure look like blurry clusters (Fig. 5b,c). In addition, inspection of the axially color-coded lateral images and axial cross-section revealed that the fibril structures in the axial direction were distorted and flattened without DL-AO. A similar phenomenon was observed in the presence of spherical aberrations in the previous evaluation of mitochondrial membranes (Figs. 3, 4 and 5b,c). Interestingly, with DL-AO, our reconstructed super-resolution images using in vitro or in situ PSF models revealed highly similar results, suggesting that DL-AO restored the aberrated emission patterns approaching the instrument optimum. Combining DL-AO with INSPR, we imaged fibril structures in different plaque areas (Fig. 5d–I), and we were able to consistently resolve individual fibrils and revealed their 3D arrangements within plaques at various stages (Fig. 5f–I). Measuring the width of Aβ fibrils in tissues, we obtained an averaged width of about 52 ± 9 nm (mean ± s.d., N = 30) and 72 ± 19 nm (mean ± s.d., N = 30) in lateral and axial cross-sections, respectively (Fig. 5j). We note that these measured fibril widths have slight variations among different imaged plaques.
Dendritic spines in 150–250-µm-cut mouse brain sections
Using DL-AO to correct sample-induced aberrations, and in situ PSF models to perform super-resolution reconstruction post-AO correction, we performed SMLM imaging through 150–250-μm-cut brain sections resolving dendritic spines, the 300–800-nm tiny protrusions from the dendrites whose morphology changes in response to neuronal activities associated with learning and memory56,57. Insufficient spatial resolution leads to an erroneous classification of spines58,59 due to their miniature sizes. The capacity to resolve spines’ ultrastructure within their tissue environment is critical in detecting morphological changes in the same area of the functional measurements. This technological advancement will allow electrophysiological and morphological mapping of the same neural circuits linking functional and structural synaptic plasticity with animal behavior60. We imaged Thy1-ChR2-EYFP transgenic mice, expressing Channelrhodopsin-2 enhanced yellow fluorescent protein (EYFP) fusion protein in cortical L5 Thy1+ pyramidal cells61. Through a 250-μm-cut brain section, we resolved the distinct membrane distribution of the fluorescently tagged target decorating the dendritic spines (Fig. 6 and Extended Data Fig. 10). Throughout the resolved volume of spines, we could observe the membrane-bounded structures as hollow tubes and blobs (Fig. 6d). Besides, the very thin neck of spines can be clearly visualized (Fig. 6g and Extended Data Fig. 10), which provides more accurate information about the dimension of spines. We also imaged 150-μm-cut mouse brain sections (Fig. 6c,f), where thinner sections provide a better signal-to-background ratio. Interestingly, we observed a few occurrences where dendrite membranes labeled ChR2-EYFP appeared to be twisted in the final reconstructed images (Fig. 6f), which may represent a type of physical substrate for decreasing gain for synaptic inputs62,63. We obtained an average localization precision of 13 nm and 57 nm in lateral and axial dimensions when imaging through the 250-μm-cut brain section, and 11–52 nm (lateral–axial) precision when imaging through the 150-μm-cut brain section. The capacity to resolve and accurately quantify the shape and size of dendritic spines through large tissue depths paves the way to link spine morphology and function and will facilitate studies of learning, memory and brain disorders.
Discussion
Combing the power of single-molecule DNN with careful designs in network training, feedback and instrument control, we demonstrated that DL-AO optimizes PSFs approaching the instrument optimum during SMLM experiments, and restores the resolution of 3D SMLM through a depth of >130 µm in brain tissues. DL-AO is demonstrated to work robustly in various types of data and specimens, including simulated SMLM frames (Fig. 2e, Supplementary Figs. 3a,b and 5), fluorescence beads (Figs. SS1 and SS2), mitochondrial networks in cells8,40 (Figs. 3 and 4 and Extended Data Figs. 8 and 9), Aβ plaques64 in the brains of mouse models of Alzheimer’ disease (Fig. 5, Supplementary Fig. 6 and Fig. SS23), as well as dendrites and spines40,65,66 in cortical L5 Thy1+ pyramidal cells in the brains of Thy1-ChR2-EYFP transgenic mice (Fig. 6 and Extended Data Fig. 10). For these data acquired at an imaging depth of 35–134 μm, a lateral resolution of 14–31 nm and a 3D resolution of 41–81 nm on average were measured using decorrelation analysis67 and Fourier shell correlation68, respectively (Supplementary Fig. 8). Throughout all these demonstrations, we have kept the DL-AO network parameters unchanged including architecture and training range. The key to the consistent performances despite the distinct sample variations lies in the detection of single molecules because these emission patterns bear no influences from the underlying structures and thus provide a unique and pure source for aberration measurements, invariant across sample types.
However, DL-AO requires at least two isolated and detectable PSFs to start compensation, and this requirement might be challenging to meet when the aberration level or imaging depth is drastically higher than the demonstrated cases where single-molecule emissions are no longer identifiable. In those cases, an initial compensation with the conventional metric-based AO method would serve as a good start while DL-AO provides subsequent and continuous fine aberration corrections for high-resolution single-molecule reconstruction. Because measuring aberrations from single-molecule-containing subregions bears no influences from the underlying sample structures, DL-AO is capable of robustly compensating aberrations despite the dynamic changes in the underlying sample structure (Supplementary Video 8). We further demonstrated that the improved compensation speed (Supplementary Videos 4 and 5) makes DL-AO capable of monitoring and compensating for random and sudden aberration changes (Supplementary Videos 6 and 7). Some of these cellular and tissue structures have been shown previously in thinner sections or on coverslip surfaces8,69,70,71. Imaging these well-characterized structures helps us in identifying the potential artifact and provides visual assessments of the achievable resolution through the complex tissue and cell environments tested here.
Further, we performed an initial investigation on controlling 50 mirror modes simultaneously with DL-AO (Supplementary Note 8). We observed that DL-AO network responded toward individual mirror deformations mostly in a one-to-one manner, a behavior observed with both beads samples and blinking single molecules from immunofluorescence-labeled cell specimens (Figs. SS19 and 20). We expect that future development in designing training data and neural network architecture will improve the inference accuracy of DL-AO through a large compensation range, ultimately enabling single-shot compensation during SMLM imaging. Additionally, the demonstrated DL-AO applications are limited by the working distance of the silicone oil objective, and thus the imaging depth could potentially be extended when combined with long working distance objectives, if permissible by tissue scattering and fluorescence background. Besides, the current implementation of DL-AO only corrects aberration shared within the field of view, because a deformable mirror is placed at the common pupil plane of the entire FOV. For the residual wavefront differences, analytical methods, such as INSPR40, can be applied to retrieve region-specific PSF models to localize molecules at different segments of the field of view (Fig. SS22). To compensate field-of-view-dependent aberrations, DL-AO could be potentially combined with the multi-pupil adaptive optics approach72. To further improve the achievable resolution and imaging fidelity, we expect that DL-AO can be combined with light-sheet illumination73,74 for an increased signal-to-background ratio of single-molecule detections, tissue clearing75 for labeling penetration and reduced aberration level and expansion methods76 for further improved spatial resolution, opening doors to observe ultrastructural organizations and colocalizations in tissues and small animals.
Finally, we note a very exciting demonstration in the previous AO-SMLM work by Siemons et al. (‘REALM’)33, which demonstrates the possibility of correlating AO-SMLM and functional measurement in brain sections. Although SMLM experiments were performed in fixed tissue, the possibility of accessing tissue nanoscale features in the context of its function illustrated an impactful direction of SMLM in neuroscience. We expect that the demonstrated capacity of DL-AO makes it a central player in connecting our understanding of the brain’s ultrastructure and function. SMLM in live tissue, however, has major challenges. Tissue-induced aberration and scattering, the limited temporal resolution, live-tissue compatible probes and its labeling strategy represent barriers in revealing the ultrastructural dynamics in living tissues and animals. DL-AO allows robust compensation of complex wavefront through tissues in near real-time. We believe it represents one solid step toward this grand challenge of live-tissue nanoscopy.
Methods
Preparation of fluorescent beads on coverslips
We cleaned 25-mm-diameter coverslips (CSHP-No1.5-25, Bioscience Tools) successively in ethanol (2701, Decon) and HPLC-grade water (W5-4, Fisher Chemical) three times and then dried them with compressed air. To promote fluorescent bead adhesion on the coverslip, 200 µl of poly-l-lysine solution (P4707, Sigma-Aldrich) was added to one coverslip and incubated for 20 min at room temperature (RT). Following that, the coverslip was rinsed with deionized water. For bead incubation, we first diluted 100-nm-diameter crimson beads (custom-designed, Invitrogen) to 1:1,000,000 in deionized water. Then we added 200 µl of the diluted bead solution to the center of the coverslip and incubated for 20 min at RT. The coverslip was subsequently rinsed with deionized water. The treated coverslip was placed on a custom-made holder13, and 20 µl of 38% 2,2′-thiodiethanol (166782, Sigma-Aldrich) in 1× PBS (10010023, Gibco) was added to its center. Another 25-mm-diameter coverslip (cleaned using the above protocol) was placed on top of this coverslip. This coverslip sandwich was sealed with two-component silicone dental glue (Twinsil speed 22, Dental-Produktions und Vertriebs).
Cell culture
COS-7 cells (CRL-1651, American Type Culture Collection (ATCC)) were grown on coverslips placed in six-well plates and cultured in DMEM (30-2002, ATCC) with 10% FBS (30-2020, ATCC) and 1% penicillin–streptomycin (15140122, Gibco) at 37 °C with 5% CO2. The cells were passaged when their confluence reached 80%. The cells were fixed when their confluence reached about 30%.
Fixation and labeling of Tom20 in COS-7 cells
Cultured cells were first fixed with 37 °C pre-warmed 3% formaldehyde aqueous solution (Formalin) diluted in 1× PBS from 16% formalin (15710, Electron Microscopy Sciences (EMS)) and 0.5% glutaraldehyde aqueous solution (diluted in 1× PBS from 8% glutaraldehyde aqueous solution, 16019, EMS), with gently rocking at RT for 15 min. After fixation, cells were rinsed twice with 1× PBS and then quenched for 7 min with freshly prepared 0.1% sodium borohydride (452882, Sigma-Aldrich) in 1× PBS. The cells were rinsed three times with 1× PBS and blocked with 3% BSA (001-000-162, Jackson ImmunoResearch) and 0.2% Triton X-100 in 1× PBS, with gently rocking at RT for 1 h. After blocking, the cells were incubated at 4 °C overnight with primary antibody (sc-11415, Santa Cruz Biotechnology), diluted at 1:500 in antibody dilution buffer (1% BSA and 0.2% Triton X-100 in 1× PBS). We then washed cells three times for 5 min each time in 0.05% Triton X-100 in 1× PBS, and incubated cells at RT for 5 h with goat anti-rabbit IgG (H + L), Alexa Fluor 647-conjugated secondary antibody (A21245, Invitrogen), diluted at 1:500 in antibody dilution buffer (1% BSA and 0.2% Triton X-100 in 1× PBS). After being washed three times with 5 min each time in 0.05% Triton X-100 in 1× PBS, cells were post-fixed with 4% formalin (diluted at 1:4 with 1× PBS from 16% formalin, 15710, EMS) at RT for 10 min. Cells were then rinsed three times with 1× PBS and stored in 1× PBS at 4 °C.
Fixation and labeling of amyloid-β in mouse brain sections
The 5xFAD Alzheimer’s disease mouse model was used for immunostaining Aβ. Mice were maintained on the C57BL/6J (B6) background (strain 000664), which were purchased from the Jackson Laboratory (JAX MMRRC, 034848). The 5xFAD transgenic mice overexpress five familial Alzheimer’s disease (FAD) mutations under control of the Thy1 promoter: the APP (695) transgene containing the Swedish (p.Lys670Asn, p.Met671Leu), Florida (p.Ile716Val) and London (p.Val717Ile) mutations, and the PSEN1 transgene containing the p.Met146Leu and p.Leu286Val FAD mutations32.
Up to five mice were housed per cage with SaniChip bedding and LabDiet 5K52/5K67 (6% fat) feed, with 40–60% humidity at 20–26 °C. The colony room was kept on a 12:12-h light–dark schedule with the lights on from 7:00 to 19:00 daily. The mice were bred and housed in specific-pathogen-free conditions.
Female mice were euthanized by perfusion with ice-cold PBS following full anesthetization with Avertin (125–250 mg per kg body weight intraperitoneal injection)77. Animals used in the study were housed in the Stark Neurosciences Research Institute Laboratory Animal Resource Center, Indiana University School of Medicine. All animals were maintained and experiments performed in accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocol was approved by the Institutional Animal Care and Use Committee at Indiana University School of Medicine.
Perfused brains from mice at 7.5 months of age were fixed in 4% formalin (1:4 dilution with 1× PBS from 16% formalin, 15710, EMS) for 24 h at 4 °C. Following fixation, brains were cryoprotected in 30% sucrose at 4 °C, and then cut into sections of 150 μm by a vibratome (7000smz-2, Campden Instruments). For immunostaining, free-floating sections were washed and permeabilized with 0.1% Triton X-100 in 1× PBS (PBST), and antigen retrieval was subsequently performed using 1× Reveal Decloaker (Biocare Medical) at 85 °C for 10 min. Sections were blocked in 5% normal donkey serum (D9663, Sigma-Aldrich) in PBST for 1 h at RT. The sections were then incubated with Aβ antibody (Cell Signaling Technology, 2454, rabbit) at a 1:1,000 dilution in 5% normal donkey serum in PBST at 4 °C overnight. Sections were washed and stained for 1 h at RT with donkey anti-rabbit IgG (H + L), Alexa Fluor 647-conjugated secondary antibody (A31573, Invitrogen) diluted at 1:1,000 in 5% normal donkey serum in PBST78.
Fixation and labeling of Thy1+ pyramid cells in mouse brain sections
To obtain mice expressing the proper amount of ChR2-EYFP in Thy1+ pyramidal cells, the litters of Thy1-ChR2-EYFP (B6.Cg-Tg (Thy1-COP4/EYFP)18Gfng/J, Jackson Laboratory) mice crossed with B6 (C57BL/6, Jackson Lab) mice were used for the labeling (mouse strain 000664; mouse species: Mus musculus). The humidity for mouse housing is 44%, and the temperature is 22 °C. The colony room was kept on a 12/12-h light–dark cycle with the light on from 6:00 to 18:00 daily.
To extract the brains for sectioning, the litters of 7-week-old mice were first anesthetized by intraperitoneal injections of a mix of 90 mg per kg body weight ketamine (59399-114-10, Akron) and 10 mg per kg body weight xylazine (343750, HVS). After confirmation of deep anesthesia, the abdomen was open to expose the diaphragm. The chest cavity was then opened by cutting through the diaphragm and ribs to expose the heart. The trans-cardiac perfusion was performed by inserting a needle into the left ventricle and a small incision into the right atrium. Mice were perfused with 1× PBS (1:10 dilution from DSP32060, Dot Scientific). After the liver was pale, mice were continuously perfused with 4% formalin (1:8 dilution with 1× PBS from 32% formalin, 15714, EMS) to pre-fix the brain until the muscle turned stiff. Brains were carefully collected and post-fixed with 4% formalin at 4 °C overnight. The fixed brains were trimmed for coronal slicing. The trimmed brains were fixed and cut into sections of 150 μm, 200 μm and 250 μm by a vibratome (1000 Plus, TPI Vibratome).
The brain sections were washed three times, for 15 min each time, in wash buffer (0.1% Triton X-100 in 1× PBS) with a gentle shake (120 r.p.m., Orbi-Shaker, Benchmark), and then were incubated in blocking butter (5% BSA (A9647, Sigma-Aldrich) in 1× PBS) for 1.5 h with a gentle shake. The blocked brain sections were incubated with chicken anti-GFP antibody (ab13970, Abcam; diluted to 1:1,000 in blocking buffer) at 4 °C overnight. After washing three times in the wash buffer as in the first step, the slices were incubated with goat anti-chicken IgY (H + L), Alexa Fluor 647-conjugated antibody (A21449, Invitrogen; diluted to 1:600 in wash buffer) at RT for 2 h with gentle rocking.
All animals were maintained and experiments performed in accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocol was approved by the Institutional Animal Care and Use Committee at Purdue University.
Imaging buffer and sample mounting for single-molecule localization microscopy
Immediately before SMLM imaging, the coverslip with specimens was placed on a custom-made holder13. Imaging buffer79 (10% (wt/vol) glucose in 50 mM Tris, 50 mM sodium chloride, 10 mM 2-mercaptoethylamine, 50 mM 2-mercaptoethanol, 2 mM cyclooctatetraene, 2.5 mM protocatechuic acid and 50 nM protocatechuic dioxygenase, pH 8.0) was added to the coverslip. Then, another cleaned coverslip was placed on top. This coverslip sandwich was sealed with two-component silicone dental glue. Samples with immunofluorescence-labeled cells on the top coverslips were prepared as described below: 200 µl of poly-l-lysine solution was added to the bottom coverslip, incubated for 20 min and subsequently rinsed with deionized water. Then, 20 µl of microsphere suspension (134 µm in diameter, 7640A, Thermo Scientific) was spread around the outer ring area of the coverslip, and incubated at RT until the coverslip was dried. Then, we placed this coverslip with microspheres at the bottom, added 50–80 µl imaging buffer without touching the microspheres, and added the coverslip with cells on top, with the cell-side surface facing down. The refractive indices of sample media and immersion oil were 1.35 and 1.406, respectively, measured by Abbe refractometer (334610, Thermo Scientific).
Microscope setup
All experimental data were recorded on a custom-designed SMLM setup built around an Olympus IX-73 microscope stand (Olympus America). This system is equipped with a ×100/1.35-NA (numerical aperture) silicone oil-immersion objective lens (UPLSAPO100XS, Olympus America), a PIFOC objective positioner (ND72Z2LAQ, Physik Instrumente), a three-axis piezo nano-positioning system (Nano-LP200, Mad City Labs) and a manual XY stage (MicroStage-LT, Mad City Labs). A continuous-wave laser at a wavelength of 642 nm (2RU-VFL-P-2000-642-B1R, MPB Communications) was coupled with a polarization-maintaining single-mode fiber (PM-S405-XP, Thorlabs) after passing through an acousto-optic tunable filter (AOTFnC-400.650-TN, AA Opto-electronic) for power modulation. The excitation light coming out of the fiber was focused onto the pupil plane of the objective lens after passing through a filter cube holding a quadband dichroic mirror (Di03-R405/488/561/635-t1, Semrock). The emission fluorescence was split with a 50/50 non-polarizing beam splitter (BS016, Thorlabs) mounted on a kinematic base (KB25/M, Thorlabs). The separated fluorescence signals were delivered by two mirrors onto a 90° specialty mirror (47-005, Edmund Optics), passed through a band-pass filter (FF01-731/137-25), and were then projected on an sCMOS camera (Orca-Flash4.0v3, Hamamatsu) with an effective pixel size of 119 nm on the sample plane. The detection planes that received the signals transmitted and reflected by the beam splitter were referred to as plane 1 and plane 2, respectively. The pupil plane of the objective lens was imaged onto a deformable mirror (Multi-3.5, Boston Micromachines). The imaging system was controlled by a custom-written program in LabVIEW (National Instruments).
Measurement of mirror deformation modes
The experimental mirror deformation modes47 (Supplementary Note 1) were measured using the fluorescence bead sample described above. We introduced positive and negative (unit amplitude) mirror changes for each mirror deformation mode. For each mirror shape setting, we acquired PSFs at z-positions from –1.5 to 1.5 µm, with a step size of 100 nm, a frame rate of 10 Hz and three frames per z-position. Pupil phase was extracted through a phase retrieval algorithm. To obtain the experimental mirror deformation bases without the influences of instrument-induced or sample-induced aberrations, we calculated the differences of the retrieved pupil phases between the positive and negative unit changes of mirror modes and divided them by two. The actual distortion level introduced by each experimental mirror mode was quantified through root-mean-square wavefront error48 (Methods and Supplementary Note 3).
Measurement of instrument optimum
We define instrument optimum as the status where optical hardware was optimized to limit the inherent system aberrations. To obtain this optimized status, we followed a previously described method6, where the deformable mirror was adjusted as follows. Starting from the flat voltage map (provided by the manufacturer) of the deformable mirror, 28 mirror modes (Fig. SS6) were applied sequentially. For each mirror mode, 11 different amplitudes were applied while recording the corresponding fluorescence signal from an in-focus 100-nm crimson bead sample. To extract the fluorescence signal from individual beads, the symmetry center of each imaged bead was obtained using the radial symmetry method80. Subsequently, a symmetric two-dimensional Gaussian was generated at the symmetry center and was multiplied by the isolated emission pattern from the fluorescence bead, generating a Gaussian-masked image, and then the total intensity of the masked image was calculated to extract the center peak signal of the beads in focus. For each mirror mode, images of the bead were acquired at 11 different mirror mode amplitudes and the corresponding center peak signals of the bead were extracted as described above. The optimal amplitude (that is, the amplitude providing the highest center peak signal from the beads) was determined from a quadratic fit of these 11 signal measurements versus mirror mode amplitudes. After identifying optimal amplitudes for each of the 28 modes, these amplitudes were added to the flat voltage map (provided by the manufacturer), serving as the starting point for another iteration. This iterative process was repeated five times to achieve optimal system aberration correction. PSFs under instrument optimum were measured using the fluorescence beads sample described above. Data were acquired at a series of z-positions from –1.5 to 1.5 µm, with a step size of 100 nm, a frame rate of 10 Hz and three frames per z-position. A phase retrieval algorithm was then performed on the bead stack to obtain the pupil function under instrument optimum. The instrument optimum can be further verified by decomposing the pupil phase into Zernike mode48 and checking whether the absolute values of the first 64 Zernike coefficients (Wyant order48) are smaller than 0.2 \(\lambda /2\pi\).
Calculation of mean square wavefront error
The root-mean-square wavefront error (\({W}_{{rms}}\)) values were calculated by the root mean square among all pixels within the image of pupil phase angle. \({W}_{{rms}}\) values for experimental wavefronts were either calculated using the pupil phase obtained by phase retrieval from fluorescence beads (Fig. SS6) or calculated using the wavefront images composed of a linear combination of experimental mirror deformation modes as estimated by DL-AO (Fig. 2e,f, Extended Data Figs. 2, 4, 5 and 7 and Supplementary Figs. 3, 5 and 7).
Measurement of network responses to individual mirror deformation modes
The aberrated PSFs for characterizing network responses (Extended Data Figs. 2 and 4 and Supplementary Fig. 1) were measured using either Tom20 specimens or fluorescence bead samples described above. The samples were first excited with the 642-nm laser at a low intensity of ~50 W/cm2 to find regions of interest. Then, data containing single-molecule blinking events were collected at a laser intensity of 2–6 kW/cm2 and a frame rate of 50 Hz. The aberrated PSFs from the fluorescence bead samples were measured the same way as we measured PSFs under the instrument optimum. A set of PSF measurements was performed under positive and negative unit changes of each mirror deformation mode. The differences of network output between positive and negative mirror changes were calculated and divided by two giving the final response vector for each mirror deformation mode.
Single-molecule localization microscopy acquisition with deep learning-driven adaptive optics
In SMLM data acquisition, the fluorescently labeled samples were first excited with a 642-nm laser at a low intensity of ~50 W/cm2 to find a region of interest. Imaging depths of mitochondrial specimens were measured by the differences of PIFOC readings between the apparent focus of the region of interest and the bottom coverslip surface. The imaging depths for immunofluorescence-labeled tissue specimens were measured by the differences of PIFOC readings between apparent focus signals of the region of interest and the fluorescence signal closest to bottom coverslip surface. The optically measured tissue thicknesses vary among samples that contain brain sections of the same machine-cut thickness. This mismatch between machine-cut thickness (for example, 250-um-cut brain sections mentioned above) and optically measured thickness might be caused by variations in media volume between bottom and top coverslips. Before SMLM experiments, bright-field images of this region were recorded over an axial range from −1 to 1 μm with a step size of 100 nm as reference images for focus stabilization81. Then, the blinking data were collected at a laser intensity of 2–6 kW/cm2 and a frame rate of 50 Hz, where the first \(\sim\)3–20 cycles were used for DL-AO, with 20–100 frames per cycle. In the case where high levels of background photons were observed (\(\sim\)100 per pixel per frame), a temporal median filter was used to estimate structured background for each pixel. This background map was subtracted from each camera frame before the frames were segmented into subregions for DL-AO processing. After DL-AO correction, 2,000 frames were collected per cycle, and 20–120 cycles (50,000–236,000 frames; Supplementary Table 1) were collected for each imaging area. For the interleaved SMLM imaging without and with AO, deformable mirror shape was set to switch between the DL-AO compensated shape and the shape used for instrument optimum (Methods) for each imaging cycle (2,000 frames). Acquisition of no-AO data was performed first in the interleaved sequence for fair comparison. Upon each switch between no-AO and DL-AO acquisitions, the PIFOC objective positioner was moved to compensate apparent focal shift in the case of index mismatch-induced aberration82. The focal shifts were determined by an estimated linear relationship between the apparent focus shift and the amplitudes of two radially symmetric mirror deformation modes. The shifts per unit amplitude changes were empirically estimated to be −0.3 µm for mirror mode 5 and −0.2 µm for mirror mode 15 (Fig. SS6). Here, a negative movement of the PIFOC objective positioner corresponds to shifting the imaging plane closer to the bottom coverslip surface.
Structure size quantification in the reconstructed images
The neck sizes of dendritic spines are measured as follows. First, we selected a profile line at the location where measurement was to be made. A rectangular box was then cropped along the line, with its width ranging from 50 to 500 nm (depending on the spine neck length and the number of localizations). The localization result inside this rectangular box was isolated and rendered into an image with a 3-nm pixel size. Each point in the rendered image was blurred with a Gaussian kernel of 3 pixels in width. Intensity profile was generated along the profile line by sum projection and subsequently the histogram was normalized by dividing its maximum value. The spine neck/head sizes were calculated by the full width at the half maximum of the intensity histogram. The widths of the Aβ fibrils were measured the same way as that for the spine necks, except for a Gaussian function was used to fit the line profile (‘fit’, Curve Fitting Toolbox 2020a, MATLAB R2020a, MathWorks), with Gaussian function switched between ‘gauss1’ (single Gaussian fit) and ‘gauss2’ (two Gaussians) depending on the number of peaks observed in the intensity histogram. The half width at the half maximum of the fitted Gaussian curve was treated as the width of each fibril.
Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.
Data availability
The results of molecular localizations for cell/tissue structures are available in Figshare at https://doi.org/10.6084/m9.figshare.23823438. Example training and testing data for DL-AO are available in Supplementary Software packages. Complete training and testing datasets can be generated through the shared codes. Other data that support the findings of this study are available from the corresponding authors upon request.
Code availability
PyTorch scripts for training, validating and testing DL-AO and MATLAB codes for generating single-molecule training datasets are available as Supplementary Software, and further updates will be made available in the GitHub repository. We also include a Jupyter Notebook in the Supplementary Software for demonstrating one-time wavefront inference using a trained DL-AO network. The code in this notebook is executable in a web browser with Google Colaboratory, which provides free access to essential computing hardware such as GPUs or TPUs used in DL-AO. Custom code used in this work is available at https://github.com/HuanglabPurdue/DL-AO.
References
Hell, S. W. Far-field optical nanoscopy. Science 316, 1153–1158 (2007).
Hell, S. W. & Wichmann, J. Breaking the diffraction resolution limit by stimulated emission: stimulated-emission-depletion fluorescence microscopy. Opt. Lett. 19, 780–782 (1994).
Wu, Y. & Shroff, H. Faster, sharper, and deeper: structured illumination microscopy for biological imaging. Nat. Methods 15, 1011–1019 (2018).
Betzig, E. et al. Imaging intracellular fluorescent proteins at nanometer resolution. Science 313, 1642–1645 (2006).
Hess, S. T., Gould, T. J., Gunewardene, M., Bewersdorf, J. & Mason, M. D. Ultrahigh resolution imaging of biomolecules by fluorescence photoactivation localization microscopy. Methods Mol. Biol. 544, 483–522 (2009).
Rust, M. J., Bates, M. & Zhuang, X. Sub-diffraction-limit imaging by stochastic optical reconstruction microscopy (STORM). Nat. Methods 3, 793–795 (2006).
Schmidt, R. et al. Spherical nanosized focal spot unravels the interior of cells. Nat. Methods 5, 539–544 (2008).
Huang, B., Wang, W., Bates, M. & Zhuang, X. Three-dimensional super-resolution imaging by stochastic optical reconstruction microscopy. Science 319, 810–813 (2008).
Xu, K., Zhong, G. & Zhuang, X. Actin, spectrin, and associated proteins form a periodic cytoskeletal structure in axons. Science 339, 452–456 (2013).
Eilers, Y., Ta, H., Gwosch, K. C., Balzarotti, F. & Hell, S. W. MINFLUX monitors rapid molecular jumps with superior spatiotemporal resolution. Proc. Natl Acad. Sci. USA 115, 6117–6122 (2018).
Moerner, W. E., Shechtman, Y. & Wang, Q. Single-molecule spectroscopy and imaging over the decades. Faraday Discuss. 184, 9–36 (2015).
Von Diezmann, A., Shechtman, Y. & Moerner, W. E. Three-dimensional localization of single molecules for super-resolution imaging and single-particle tracking. Chem. Rev. 117, 7244–7275 (2017).
Huang, F. et al. Ultra-high resolution 3D imaging of whole cells. Cell 166, 1028–1040 (2016).
Liu, S. & Huang, F. Enhanced 4Pi single-molecule localization microscopy with coherent pupil based localization. Commun. Biol. 3, 220 (2020).
Bates, M. et al. Optimal precision and accuracy in 4Pi-STORM using dynamic spline PSF models. Nat. Methods 19, 603–612 (2022).
Oi, C. et al. LIVE-PAINT allows super-resolution microscopy inside living cells using reversible peptide–protein interactions. Commun. Biol. 3, 458 (2020).
Smith, C. S., Joseph, N., Rieger, B. & Lidke, K. A. Fast, single-molecule localization that achieves theoretically minimum uncertainty. Nat. Methods 7, 373–375 (2010).
Liu, S., Huh, H., Lee, S.-H. & Huang, F. Three-dimensional single-molecule localization microscopy in whole-cell and tissue specimens. Annu. Rev. Biomed. Eng. 22, 155–184 (2020).
Tyson, R. Principles of Adaptive Optics (Academic Press, 1991).
Hardy, J. W. Adaptive Optics for Astronomical Telescopes (Oxford University Press, 1998).
Booth, M., Andrade, D., Burke, D., Patton, B. & Zurauskas, M. Aberrations and adaptive optics in super-resolution microscopy. Microscopy 64, 251–261 (2015).
Ji, N. Adaptive optical fluorescence microscopy. Nat. Methods 14, 374–380 (2017).
Hampson, K. M. et al. Adaptive optics for high-resolution imaging. Nat. Rev. Methods Primers 1, 68 (2021).
Wang, K. et al. Rapid adaptive optical recovery of optimal resolution over large volumes. Nat. Methods 11, 625–628 (2014).
Liu, R., Li, Z., Marvin, J. S. & Kleinfeld, D. Direct wavefront sensing enables functional imaging of infragranular axons and spines. Nat. Methods 16, 615–618 (2019).
Hu, L., Hu, S., Gong, W. & Si, K. Deep learning assisted Shack–Hartmann wavefront sensor for direct wavefront detection. Opt. Lett. 45, 3741–3744 (2020).
Turcotte, R. et al. Dynamic super-resolution structured illumination imaging in the living brain. Proc. Natl Acad. Sci. USA 116, 9586–9591 (2019).
Li, Z. et al. Fast widefield imaging of neuronal structure and function with optical sectioning in vivo. Sci. Adv. 6, eaaz3870 (2020).
Burke, D., Patton, B., Huang, F., Bewersdorf, J. & Booth, M. J. Adaptive optics correction of specimen-induced aberrations in single-molecule switching microscopy. Optica 2, 177–185 (2015).
Tehrani, K. F., Xu, J., Zhang, Y., Shen, P. & Kner, P. Adaptive optics stochastic optical reconstruction microscopy (AO-STORM) using a genetic algorithm. Opt. Express 23, 13677–13692 (2015).
Tehrani, K. F., Zhang, Y., Shen, P. & Kner, P. Adaptive optics stochastic optical reconstruction microscopy (AO-STORM) by particle swarm optimization. Biomed. Opt. Express 8, 5087–5097 (2017).
Mlodzianoski, M. J. et al. Active PSF shaping and adaptive optics enable volumetric localization microscopy through brain sections. Nat. Methods 15, 583–586 (2018).
Siemons, M. E., Hanemaaijer, N. A. K., Kole, M. H. P. & Kapitein, L. C. Robust adaptive optics for localization microscopy deep in complex tissue. Nat. Commun. 12, 3407 (2021).
Booth, M. J., Neil, M. A. A., Juskaitis, R. & Wilson, T. Adaptive aberration correction in a confocal microscope. Proc. Natl Acad. Sci. USA 99, 5788–5792 (2002).
Débarre, D., Booth, M. J. & Wilson, T. Image based adaptive optics through optimisation of low spatial frequencies. Opt. Express 15, 5131–5139 (1991).
Žurauskas, M. et al. IsoSense: frequency enhanced sensorless adaptive optics through structured illumination. Optica 6, 370–379 (2019).
Thomas, B., Wolstenholme, A., Chaudhari, S. N., Kipreos, E. T. & Kner, P. Enhanced resolution through thick tissue with structured illumination and adaptive optics. J. Biomed. Opt. 20, 26006 (2015).
Débarre, D., Botcherby, E. J., Booth, M. J. & Wilson, T. Adaptive optics for structured illumination microscopy. Opt. Exp. 16, 9290–9305 (2008).
Gould, T. J., Burke, D., Bewersdorf, J. & Booth, M. J. Adaptive optics enables 3D STED microscopy in aberrating specimens. Opt. Express 20, 20998 (2012).
Xu, F. et al. Three-dimensional nanoscopy of whole cells and tissues with in situ point spread function retrieval. Nat. Methods 17, 531–540 (2020).
Zhang, P. et al. Analyzing complex single-molecule emission patterns with deep learning. Nat. Methods 15, 913–916 (2018).
LeCun, Y. et al. Backpropagation applied to digit recognition. Neural Comput. 1, 541–551 (1989).
LeCun, Y., Bottou, L., Bengio, Y. & Haffner, P. Gradient-based learning applied to document recognition. Proc. IEEE 86, 2278–2323 (1998).
He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. in Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition 770–778 (2016).
He, K., Zhang, X., Ren, S. & Sun, J. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. in Proceedings of the IEEE International Conference on Computer Vision 1026–1034 (2015).
Ioffe, S. & Szegedy, C. Batch normalization: accelerating deep network training by reducing internal covariate shift. in Proceedings of the 32nd International Conference on Machine Learning 448–456 (2015).
Wang, B. & Booth, M. J. Optimum deformable mirror modes for sensorless adaptive optics. Opt. Commun. 282, 4467–4474 (2009).
Wyant, J. C. & Creath, K. Basic wavefront aberration theory for optical metrology. in Applied Optics and Optical Engineering Vol. 11 (Academic Press, 1992).
Kay, S. M. Fundamentals of Statistical Signal Processing, Volume I: Estimation Theory (Prentice Hall, 1993).
Hanser, B. M., Gustafsson, M. G. L., Agard, D. A. & Sedat, J. W. Phase-retrieved pupil functions in wide-field fluorescence microscopy. J. Microsc. 216, 32–48 (2004).
Haber, A. & Bifano, T. General approach to precise deformable mirror control. Opt. Express 29, 33741–33759 (2021).
Wurm, C. A. et al. Nanoscale distribution of mitochondrial import receptor Tom20 is adjusted to cellular conditions and exhibits an inner-cellular gradient. Proc. Natl Acad. Sci. USA 108, 13546–13551 (2011).
Salloway, S. et al. Amyloid-related imaging abnormalities in 2 phase 3 studies evaluating aducanumab in patients with early Alzheimer disease. JAMA Neurol. 79, 13–21 (2022).
Sevigny, J. et al. The antibody aducanumab reduces Aβ plaques in Alzheimer’s disease. Nature 537, 50–56 (2016).
Oakley, H. et al. Intraneuronal β-amyloid aggregates, neurodegeneration, and neuron loss in transgenic mice with five familial Alzheimer’s disease mutations: potential factors in amyloid plaque formation. J. Neurosci. 26, 10129–10140 (2006).
Berry, K. P. & Nedivi, E. Spine dynamics: are they all the same? Neuron 96, 43–55 (2017).
Pchitskaya, E. & Bezprozvanny, I. Dendritic spines shape analysis—classification or clusterization? Perspective. Front. Synaptic Neurosci. 12, 31 (2020).
Tønnesen, J., Katona, G., Rózsa, B. & Nägerl, U. V. Spine neck plasticity regulates compartmentalization of synapses. Nat. Neurosci. 17, 678–685 (2014).
Runge, K., Cardoso, C. & de Chevigny, A. Dendritic spine plasticity: function and mechanisms. Front. Synaptic Neurosci. 12, 36 (2020).
Kissinger, S. T. et al. Visual experience-dependent oscillations and underlying circuit connectivity changes are impaired in Fmr1 KO mice. Cell Rep. 31, 107486 (2020).
Arenkiel, B. R. et al. In vivo light-induced activation of neural circuitry in transgenic mice expressing channelrhodopsin-2. Neuron 54, 205–218 (2007).
Kubota, Y. et al. Conserved properties of dendritic trees in four cortical interneuron subtypes. Sci. Rep. 1, 89 (2011).
Behabadi, B. F. & Mel, B. W. J4 at Sweet 16: a new wrinkle? Neural Comput. 19, 2865–2870 (2007).
Lührs, T. et al. 3D structure of Alzheimer’s amyloid-β(1–42) fibrils. Proc. Natl Acad. Sci. USA 102, 17342–17347 (2005).
Pfeiffer, T. et al. Chronic 2P-STED imaging reveals high turnover of dendritic spines in the hippocampus in vivo. Elife 7, e34700 (2018).
Yuste, R. & Bonhoeffer, T. Genesis of dendritic spines: insights from ultrastructural and imaging studies. Nat. Rev. Neurosci. 5, 24–34 (2004).
Descloux, A., Grußmayer, K. S. & Radenovic, A. Parameter-free image resolution estimation based on decorrelation analysis. Nat. Methods 16, 918–924 (2019).
Diederich, B., Then, P., Jugler, A., Forster, R. & Heintzmann, R. cellSTORM—cost-effective super-resolution on a cellphone using dSTORM. PLoS ONE 14, e0209827 (2019).
Gu, L. et al. Molecular-scale axial localization by repetitive optical selective exposure. Nat. Methods 18, 369–373 (2021).
Dong, D. et al. Super-resolution fluorescence-assisted diffraction computational tomography reveals the three-dimensional landscape of the cellular organelle interactome. Light. Sci. Appl. 9, 11 (2020).
Zhao, W. et al. Sparse deconvolution improves the resolution of live-cell super-resolution fluorescence microscopy. Nat. Biotechnol. 40, 606–617 (2022).
Park, J.-H., Kong, L., Zhou, Y. & Cui, M. Large-field-of-view imaging by multi-pupil adaptive optics. Nat. Methods 14, 581–583 (2017).
Legant, W. R. et al. High-density three-dimensional localization microscopy across large volumes. Nat. Methods 13, 359–365 (2016).
Power, R. M. & Huisken, J. A guide to light-sheet fluorescence microscopy for multiscale imaging. Nat. Methods 14, 360–373 (2017).
Gradinaru, V., Treweek, J., Overton, K. & Deisseroth, K. Hydrogel-tissue chemistry: principles and applications. Annu. Rev. Biophys. 47, 355–376 (2018).
Chen, F., Tillberg, P. W. & Boyden, E. S. Expansion microscopy. Science 347, 543–548 (2015).
Tsai, A. P. et al. PLCG2 is associated with the inflammatory response and is induced by amyloid plaques in Alzheimer’s disease. Genome Med. 14, 17 (2022).
Tsai, A. P. et al. INPP5D expression is associated with risk for Alzheimer’s disease and induced by plaque-associated microglia. Neurobiol. Dis. 153, 105303 (2021).
Olivier, N., Keller, D., Gönczy, P. & Manley, S. Resolution doubling in 3D-STORM imaging through improved buffers. PLoS ONE 8, e69004 (2013).
Parthasarathy, R. Rapid, accurate particle tracking by calculation of radial symmetry centers. Nat. Methods 9, 724–726 (2012).
Mcgorty, R., Kamiyama, D. & Huang, B. Active microscope stabilization in three dimensions using image correlation. Opt. Nano. 2, 3 (2013).
Petrov, P. N. & Moerner, W. E. Addressing systematic errors in axial distance measurements in single-emitter localization microscopy. Opt. Express 28, 18616–18632 (2020).
Acknowledgements
We thank Fan Xu (Beijing Institute of Technology) for suggestions on PSF segmentation and the super-resolution reconstruction process. We thank Sheng Liu (The University of New Mexico) for suggestions on the phase retrieval algorithm and the PSF generation process. We thank Yue Zheng (Purdue University) for suggestions on the manuscript. This work was supported by the US National Institutes of Health (grants GM119785 to F.H., MH123401 to F.H. and A.A.C. and RF1AG074566 to G.E.L.).
Author information
Authors and Affiliations
Contributions
P.Z. and F.H. conceived the project and designed the experiments for DL-AO characterization. P.Z. developed the DL-AO workflow, wrote the DL-AO instrument control, performed experiments and analyzed the data. D.M. developed the microscope setup. P.Z., D.M. and H.-C.G. performed and optimized deformable mirror calibration. X.C. and A.P.T. optimized the staining procedure for tissue specimens. P.Z., X.C., A.P.T., Y.T., L.F., C.B., A.A.C. and F.H. designed the experiments and prepared biological samples. G.E.L., A.A.C. and F.H. supervised the study. All authors wrote the paper.
Corresponding authors
Ethics declarations
Competing interests
P.Z. and F.H. are inventors on a patent application submitted by Purdue University that covers the basic principles of DL-AO. All other authors have no competing interests.
Peer review
Peer review information
Nature Methods thanks Wei Gong and the other, anonymous, reviewer for their contribution to the peer review of this work. Primary Handling Editor: Rita Strack, in collaboration with the Nature Methods team. Peer reviewer reports are available.
Additional information
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Extended data
Extended Data Fig. 1 Inconsistent responses in metric-based AO.
(A) Simulated emission patterns from single molecules at different axial positions. (B) Metric values vs. amplitudes of mirror shapes. Each metric value was calculated through a weighted sum of the Fourier transform of an acquisition under certain mirror shape, the weighting factor of which is defined in previous work32,33. The acquisitions were simulated as images with 400 × 400 pixels and 65 nm pixel size. Each acquisition contains one PSF with 10000 photon counts and 10 background photon counts. Each row shows the changes in metric values when scanning the amplitudes of a mirror shape shown on the left. Different columns show metrics calculated from PSFs at different axial positions. ‘DAst’ and ‘Sph’ stand for Diagonal Astigmatism and Primary Spherical in Zernike polynomials48.
Extended Data Fig. 2 Characterizing neural network responses to mirror mode changes using PSFs measured from fluorescent beads.
(A) Network response to individual mirror mode changes. Each row of the response matrix shows the network responded mirror coefficients under a unit change of each mirror deformation mode. After linear combining measured mirror modes (images below the title) with network responded coefficients, we obtained network estimated wavefront shape w.r.t. individual mirror mode changes (the 2nd column). The 1st column shows phase retrieved wavefronts from beads imaged individual mirror mode changes. The PSFs were measured with 100-nm-diameter crimson beads. PSFs from −1.5 µm to 1.5 µm around the focus, with 0.1 µm step size, were collected for characterizing network responses. (B) Difference between network estimated wavefront and phase retrieved wavefront (the first two columns in A). The top row shows the pixel-wise differences between wavefronts obtained from network estimation and that obtained from phase retrieval. The plot below shows the root mean square wavefront error48 (\({W}_{{rms}}\), Methods) of each wavefront difference. (C) Similarity between network estimated wavefront and phase retrieved wavefront. The similarity is quantified with 2D normalized cross correlation (NCC).
Extended Data Fig. 3 Comparison between measured PSFs and PSFs simulated from network estimations based on a single measurement of an isolated molecule.
The left column shows measured PSFs from 100-nm-diameter crimson beads when scanning Piezo stage at different axial positions. The measured PSFs in biplane sub-regions were sent to neural network, which outputs a vector of mirror mode coefficients for each sub-region. The measured mirror modes (Supplementary Note 3) were linear combined with mirror mode coefficients output from network, which result in wavefronts shown in the middle column. The wavefront for network estimations based on all measured PSFs were generated with an averaged value among network outputs w.r.t. PSFs at different axial positions. The wavefront was then used to simulate PSFs at different axial positions to check the similarity between measured PSFs and PSFs simulated from network estimations. PSFs were simulated without background and noise for visualization. Scale bars: 2 µm.
Extended Data Fig. 4 Characterizing neural network responses to mirror mode changes using PSFs measured from blinking molecules.
(A) Network response to individual mirror mode changes. Each row of the response matrix shows the network responded mirror coefficients under a unit change of each mirror deformation mode. After linear combining measured mirror modes (images below the title) with network responded coefficients, we obtained network estimated wavefront shape w.r.t. individual mirror mode changes. The PSFs were measured experimental blinking frames from immune-fluorescence-labeled Tom20 specimen. 100 PSFs were used for calculating each network response. (B) Difference between network estimated wavefront (left column in A) and phase retrieved from beads (left column in Extended Data Fig. 2A). The top row shows the pixel-wise differences between wavefronts obtained from network estimation and that obtained from phase retrieval. The plot below shows the root mean square wavefront error48 (\({W}_{{rms}}\), Methods) of each wavefront difference. (C) Similarity between network estimated wavefront and phase retrieved wavefront. The similarity is quantified with 2D normalized cross correlation (NCC).
Extended Data Fig. 5 SMLM frames before and after DL-AO compensating various amount of induced aberrations.
DL-AO was compensating aberrations in different levels (in \({W}_{{rms}}\), Methods) based on experimental blinking frames from immune-fluorescence-labeled Tom20 specimen. 20 camera frames were used for DL-AO estimation before each mirror update. The blinking data after DL-AO were compensation results after 19 mirror updates. The results shown are representatives of 15 tests.
Extended Data Fig. 6 PSF shape before and after each mirror update during DL-AO compensation.
(A) Examples of PSFs before and after each mirror update, when compensating artificially induced aberrations with DL-AO. Compensations were performed in real time during SMLM experiments. The results shown are representatives of 11 tests. The SMLM blinking frames for compensation were acquired from immune-fluorescence-labeled Tom20 specimen. PSFs were measured from 100-nm-diameter crimson beads near the compensation area post SMLM acquisition. Scale bar: 5 µm. ‘PSF #’ stands for number of sub-regions used for each DL-AO network estimation. ‘phase’ stands for pupil phase obtained by phase retrieval on the measured PSFs from beads. Each PSF was normalized to maximum equals to 1. (B) Quantitative comparisons between PSFs measured under instrument optimum and those measured before and after each mirror update using 3D normalized cross correlation (NCC).
Extended Data Fig. 7 PSFs before and after DL-AO at various number of induced aberrations.
(A) Examples of PSFs before and after DL-AO, when compensating artificially induced aberrations. Compensations are performed in real time during SMLM experiments shown in Extended Data Fig. 5. PSFs are measured from 100-nm-diameter crimson beads near the compensation area post SMLM acquisition. Scale bar: 5 µm. (B) PSFs are measured under instrument optimum (Methods) from 100-nm-diameter crimson beads. Scale bar: 5 µm. (C) Quantitative comparisons between PSFs measured under instrument optimum and those measured before and after DL-AO using 3D normalized cross correlation (NCC).
Extended Data Fig. 8 Demonstrations of DL-AO correcting index mismatch induced aberration by imaging Tom20 proteins in COS-7 cells through 94 μm water-based imaging media.
(A) 3D SMLM reconstruction of Tom20 imaged through 94 μm water-based media without AO, then reconstructed with in situ PSF model (INSPR) (B) 3D SMLM reconstruction of Tom20 imaged through 94 μm water-based media with DL-AO, then reconstructed with INSPR. (C) Axial cross-sections in A and B comparing cases without and with DL-AO combined with reconstruction methods of either in vitro PSF model (PR) or in situ PSF models (INSPR). The PR PSF model for no AO case was obtained from 100-nm-diameter crimson bead next to the imaged area. The in vitro model for DL-AO was obtained from beads at the bottom coverslip surface. Choices of in vitro model are made in order to find the closest match with corresponding experimental conditions. (D) Comparison of measured PSFs at 94 μm without and with DL-AO, in situ PSF models without and with DL-AO, and the instrument optimum. Scale bar: 2 μm.
Extended Data Fig. 9 Demonstrations of DL-AO correcting index mismatch induced aberration by imaging Tom20 proteins in COS-7 cells through 35 μm water-based imaging media.
(A) 3D SMLM reconstruction of Tom20 imaged through 35 μm water-based media without AO, then reconstructed with in situ PSF model (INSPR) (B) 3D SMLM reconstruction of Tom20 imaged through 35 μm water-based media with DL-AO, then reconstructed with INSPR. (C) Axial cross-sections in A and B comparing cases without and with DL-AO combined with reconstruction methods of either in vitro PSF model (PR) or in situ PSF models (INSPR). The PR PSF model for no AO case was obtained from 100-nm-diameter crimson bead next to the imaged area. The in vitro model for DL-AO was obtained from beads at the bottom coverslip surface. Choices of in vitro model are made to find the closest match with corresponding experimental conditions. (D) Comparison of measured PSFs at 35 μm without and with DL-AO, in situ PSF models without and with DL-AO, and the instrument optimum. Scale bar: 2 μm.
Extended Data Fig. 10 Size measurements of spines’ heads and necks.
(A) Identified spines in Fig. 6A–C, and the corresponding size measurements of their necks and heads. ‘Norm. I.’ stands for normalized intensity, where intensity in reconstructed image reflects counts of localized single molecules. ‘dist.’ stands for distance. The histograms show the raw intensity counts along the lines indicated by white arrows. Sizes are measured at the full widths at the half maximum intensity (Methods). The images of spines share the same scale bar as the first image, unless labeled specifically. (B) Size measurements of spines’ heads and necks from seven dendrites in immune-fluorescence-labeled Thy1-ChR2-EYFP in 150–250 μm brain sections of 7-week-old mice. The sizes were measured from super-resolution reconstructions of Thy1-ChR2-EYFP using SMLM with DL-AO through 150-μm-cut brain sections (Dendrite 1-3) and 250-μm-cut brain sections (Dendrite 4-7). Dendrite 1-7 were reconstructed at 67 μm, 67 μm, 83 μm, 134 μm, 134 μm, 126 μm, 132 μm from coverslip surface. The imaged structures were found at depths near the axial limit of tissue thicknesses. Optically measured tissue thicknesses vary among samples, which might be caused by variations in media volume between bottom and top coverslips.
Supplementary information
Supplementary Information
Supplementary Figs. 1–9, Supplementary Videos 1–12, Supplementary Tables 1–4 and Supplementary Notes 1–9.
Supplementary Video 1
Single-molecule blinking frames during DL-AO for SMLM, when compensating artificially induced aberrations. Compensations are performed in real time during SMLM experiments when imaging immunofluorescence-labeled Tom20 specimen. The left two panels are the raw blinking frames (after converting the analog-to-digital unit readings in camera frames to the effective photoelectrons, referred as photon number) captured from the two detection planes of a biplane setup. The displayed frame rate is set to be the same as the acquisition frame rate, which is 50 Hz. The top-right panel shows the deformable mirror voltage map with respect to current detection. The deformable mirror was updated every 50 camera frames. The amount of mirror update proposed by DL-AO will be displayed on the bottom-right panel every 50 camera frames.
Supplementary Video 2
Single-molecule blinking frames during DL-AO for SMLM, when compensating aberrations induced by refractive index mismatch. Compensations are performed in real time during SMLM experiments when imaging immunofluorescence-labeled Tom20 specimen. The specimen is located at 134 μm (Methods) with refractive indices of sample media and immersion oil being 1.35 and 1.406, respectively, measured by Abbe refractometer (334610, Thermo Scientific). The left two panels are the raw blinking frames (after converting the analog-to-digital unit readings in camera frames to the effective photoelectrons, referred as photon number) captured from the two detection planes of a biplane setup. The displayed frame rate is set to be the same as the acquisition frame rate, which is 50 Hz. The top-right panel shows the deformable mirror voltage map with respect to current detection. The deformable mirror was updated every 30 camera frames. The amount of mirror update proposed by DL-AO will be displayed on the bottom-right panel every 30 camera frames. Apparent focal planes were adjusted by moving the sample stage during compensation, because index mismatch-induced aberration causes an apparent focal plane shift4.
Supplementary Video 3
Single-molecule blinking frames during DL-AO for SMLM, when compensating aberrations induced by inhomogeneous refractive indices of 101 µm mouse brain tissue. Compensations are performed in real time during SMLM experiments when imaging immunofluorescence-labeled amyloid-β fibrils of 7.5-month-old 5XFAD female mice. The left two panels are the raw blinking frames (after converting the analog-to-digital unit readings in camera frames to the effective photoelectrons, referred to as photon number) captured from the two detection planes of a biplane setup. The displayed frame rate is set to be the same as the acquisition frame rate, which is 50 Hz. The top-right panel shows the deformable mirror voltage map with respect to current detection. The deformable mirror was updated every 100 camera frames. The amount of mirror update proposed by DL-AO will be displayed on the bottom-right panel every 100 camera frames.
Supplementary Video 4
Compensation speed comparison between metric-based AO and DL-AO. Compensations are performed with in-focus PSFs from 100-nm-fluorescent beads. The left and right panel shows the raw data during the metric-based AO compensation and DL-AO compensation, respectively. The timestamp of each camera frame during imaging with AO is displayed on the top-right corner of each panel. The bottom-left panel shows the deformable mirror voltage map with respect to current detection. The deformable mirror was updated with every one camera frame. The gray levels indicate the photon counts per pixel. The timestamps were obtained from recorded files in each mirror updating cycle.
Supplementary Video 5
Comparison between metric-based AO and DL-AO on compensating aberrations when PSFs are out of focus. Compensations are performed with 100-nm-fluorescent beads that are slightly out of focus. The left and right panels show the raw data during the metric-based AO compensation and DL-AO compensation, respectively. The timestamp of each camera frame during imaging with AO is displayed on the top-right corner of each panel. The bottom-left panel shows the deformable mirror voltage map with respect to current detection. The deformable mirror was updated with every one camera frame. The gray levels indicate the photon counts per pixel. The timestamps were obtained from recorded files in each mirror updating cycle.
Supplementary Video 6
DL-AO compensates dynamic aberrations in real time. Compensations are performed in real time during SMLM experiments when imaging immunofluorescence-labeled Tom20 in COS-7 cells. The two panels in the upper row are the raw blinking frames (after converting the analog-to-digital unit readings in camera frames to the effective photoelectrons, referred to as photon number) captured from the two detection planes of a biplane setup. The wavefront shape with respect to current detection is shown at the bottom-left corner of the upper-left panel. The deformable mirror was updated every 50 camera frames. The displayed frame rate is set to be the same as the acquisition frame rate, which is 50 Hz. The number of mirror updates proposed by DL-AO after each distortion will be displayed on the bottom-left panel. The plot on the bottom row shows the distortion-level changes after each mirror update. A dot with a blue circle corresponds to a mirror update that introduces a random wavefront distortion (targeted level of 0.75 rad in \({W}_{{rms}}\)). The dots without blue circles correspond to mirror updates driven by the deep neural network. The single-molecule blinking frames with random and sudden wavefront changes were continuously acquired for 3 min and 4 min from the immunofluorescence-labeled Tom20 in COS-7 cells for Supplementary Videos 6 and 7, respectively.
Supplementary Video 7
DL-AO compensates dynamic aberrations in real time. Compensations are performed in real time during SMLM experiments when imaging immunofluorescence-labeled Tom20 in COS-7 cells. The two panels in the upper row are the raw blinking frames (after converting the analog-to-digital unit readings in camera frames to the effective photoelectrons, referred as photon number) captured from the two detection planes of a biplane setup. The wavefront shape with respect to current detection is shown at the bottom-left corner of the upper-left panel. The deformable mirror was updated every 50 camera frames. The displayed frame rate is set to be the same as the acquisition frame rate, which is 50 Hz. The number of mirror updates proposed by DL-AO after each distortion will be displayed on the bottom-left panel. The plot on the bottom row shows the distortion-level changes after each mirror update. A dot with a blue circle corresponds to a mirror update that introduces a random wavefront distortion (targeted level of 0.75 rad in \({W}_{{rms}}\)). The dots without blue circles correspond to mirror updates driven by the deep neural network. The single-molecule blinking frames with random and sudden wavefront changes were continuously acquired for 3 min and 4 min from the immunofluorescence-labeled Tom20 in COS-7 cells for Supplementary Videos 6 and 7, respectively.
Supplementary Video 8
Comparison between metric-based AO and DL-AO on compensating aberration through dynamic structure changes. The two panels are the raw blinking frames (after converting the analog-to-digital unit readings in camera frames to the effective photoelectrons, referred to as photon number) captured during metric-based AO and DL-AO, respectively. Compensations are performed in real-time during SMLM experiments when imaging immunofluorescence-labeled Tom20 in COS-7 cells. The bottom-left panel shows the deformable mirror voltage map with respect to current detection. The gray levels indicate the photon counts per pixel. The displayed frame rate is set to be the same as the acquisition frame rate, which is 50 Hz. The deformable mirror was updated every 10 camera frames when using DL-AO, while the specimen was shifted 10 nm in between each adjacent frame by the nano-positioning system (Nano-LP200, Mad City Labs) to introduce structural changes before DL-AO propose each mirror update.
Supplementary Video 9
Single-molecule blinking frames acquired with and without DL-AO when imaging through brain sections. The two panels are the raw blinking frames (after converting the analog-to-digital unit readings in camera frames to the effective photoelectrons, referred to as photon number) captured without and with DL-AO, respectively. The gray levels indicate the photon counts per pixel. The acquisition frame rate is 50 Hz. Twenty-two different imaging areas are shown in this video. These raw data were acquired from the following specimens: immunofluorescence-labeled amyloid-β plaques in mouse brain sections, immunofluorescence-labeled Thy1-ChR2-EYFP in mouse brain sections and Tom20 proteins in COS-7 cells placed on top of unlabeled mouse brain sections.
Supplementary Video 10
Single-molecule blinking frames acquired when DL-AO is compensating artificially induced aberrations in astigmatism-based setup. Compensations are performed in real time during SMLM experiments when imaging immunofluorescence-labeled Tom20 specimens. The left panel shows the raw blinking frames (after converting the analog-to-digital unit readings in camera frames to the effective photoelectrons, referred to as photon number). The displayed frame rate is set to be the same as the acquisition frame rate, which is 50 Hz. The top-right panel shows the deformable mirror voltage map with respect to current detection. The deformable mirror was updated every 100 camera frames. A background map estimated by the temporal median filter was subtracted from each camera frame before segmentation. The intensity of each subregion was estimated by summing up the photon counts in each pixel, after subtracting the median map. An intensity threshold of 1,500 photons was applied to the segmented subregions to filter out PSFs with low photon counts.
Supplementary Video 11
Single-molecule blinking frames acquired when DL-AO is compensating artificially induced aberrations in astigmatism-based setup. Compensations are performed in real time during SMLM experiments when imaging immunofluorescence-labeled Tom20 specimens. The left panel shows the raw blinking frames (after converting the analog-to-digital unit readings in camera frames to the effective photoelectrons, referred to as photon number). The displayed frame rate is set to be the same as the acquisition frame rate, which is 50 Hz. The top-right panel shows the deformable mirror voltage map with respect to current detection. The deformable mirror was updated every 100 camera frames. A background map estimated by the temporal median filter was subtracted from each camera frame before segmentation. The intensity of each subregion was estimated by summing up the photon counts in each pixel, after subtracting the median map. An intensity threshold of 1,500 photons was applied to the segmented subregions to filter out PSFs with low photon counts.
Supplementary Video 12
Single-molecule blinking frames when DL-AO is compensating aberrations induced by refractive index mismatch in astigmatism-based setup. Compensations are performed in real time during SMLM experiments when imaging immunofluorescence-labeled Tom20 specimens. The specimen locates at 146 μm (Methods) with refractive indices of sample media and immersion oil being 1.35 and 1.406, respectively, measured by Abbe refractometer (334610, Thermo Scientific). The left panel shows the raw blinking frames (after converting the analog-to-digital unit readings in camera frames to the effective photoelectrons, referred to as photon number). The displayed frame rate is set to be the same as the acquisition frame rate, which is 50 Hz. The top-right panel shows the deformable mirror voltage map with respect to current detection. The deformable mirror was updated every 100 camera frames. Apparent focal planes were adjusted by moving the sample stage during compensation, because index mismatch-induced aberration causes an apparent focal plane shift4. A background map estimated by the temporal median filter was subtracted from each camera frame before segmentation. The intensity of each subregion was estimated by summing up the photon counts in each pixel, after subtracting the median map. An intensity threshold of 1,500 photons was applied to the segmented subregions to filter out PSFs with low photon counts.
Supplementary Software
This software is distributed as accompanying software for this paper.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Zhang, P., Ma, D., Cheng, X. et al. Deep learning-driven adaptive optics for single-molecule localization microscopy. Nat Methods 20, 1748–1758 (2023). https://doi.org/10.1038/s41592-023-02029-0
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1038/s41592-023-02029-0
This article is cited by
-
Near-infrared II fluorescence imaging
Nature Reviews Methods Primers (2024)