Deep super-resolution imaging of thick tissue using structured illumination with adaptive optics

  1. Jingyu Wang
  2. Danail Stoychev
  3. Mick Phillips
  4. David Miguel Susano Pinto
  5. Richard M. Parton
  6. Nick Hall
  7. Josh Titlow
  8. Ana Rita Faria
  9. Matthew Wincott
  10. Dalia Gala
  11. Andreas Gerondopoulos
  12. Niloufer Irani
  13. Ian Dobbie
  14. Lothar Schermelleh
  15. Martin Booth
  16. Ilan Davis

Reviewed by

Read the full article

Listed in

Log in to save this article

Abstract

3D structured illumination microscopy (3D-SIM) doubles the resolution of fluorescence imaging in lateral and axial directions and increases contrast in both fixed and live specimens. However, 3D-SIM has so far not been widely applied to imaging deep in thick tissues due to its sensitivity to specimen-induced aberrations, making the method difficult to apply beyond 10 µm in depth. Furthermore, 3D-SIM has not been available in an upright configuration, limiting its use for live imaging while manipulating the specimen, for example with electrophysiology. Here, we have overcome these barriers by developing a novel upright 3D-SIM system (termed Deep3DSIM) that incorporates adaptive optics (AO) for aberration correction and remote focusing, reducing artefacts, and removing the need to move the specimen or objective. Both these advantages are equally applicable to inverted 3D-SIM microscopes. We demonstrate high-quality 3D-SIM imaging up to 130 µm into complex tissue and live sample manipulation, including human cells and Drosophila larval brains and embryos.

Article activity feed

  1. Review Commons

    Note: This response was posted by the corresponding author to Review Commons. The content has not been altered except for formatting.

    Learn more at Review Commons

    Reply to the reviewers

    Note: This preprint has been reviewed by subject experts for Review Commons. Content has not been altered except for formatting.

    Reply to the Reviewers

    We want to thank the three reviewers for their invaluable and constructive feedback. We respond to each comment individually, describing how we plan to address them in our revised manuscript.

    Reviewer #1

    1. Given the emphasis on super-resolution imaging deep inside a sample, we were surprised to see no mention of other forms of structured illumination that allow super-resolution imaging in samples thicker than a single cell. These include the 'spot-scanning' implementations of SIM that offer better imaging at depth by virtue of pinholes, and include MSIM, iSIM, and rescan confocal technologies. The two-photon / AO implementation of iSIM seems particularly germane, e.g. https://pubmed.ncbi.nlm.nih.gov/28628128/ Please consider citing these works, as they help place the existing work into context.

    Response:

    We want to thank reviewer #1 for the good point. To address this comment, we plan to add to the discussion section a description of these super resolution techniques, together with other SIM methods, explaining how they compare to our approach.

    1. As we're sure the authors appreciate, besides aberrations, a major additional obstacle to 3D SIM in thick tissues is the presence of out-of-focus background. Indeed, this point was mentioned by Gustafsson in his classic 2008 paper on 3D SIM (https://pubmed.ncbi.nlm.nih.gov/18326650/): 'The application area of three-dimensional structured illumination microscopy overlaps with that of confocal microscopy, but the two techniques have different and complementary strengths. Structured illumination microscopy offers higher effective lateral resolution, because it concentrates much of the excitation light at the very highest illumination angles, which are most effective for encoding high-resolution information into the observed data, whereas confocal microscopy spreads out its illumination light more or-less uniformly over all available angles to form a focused beam. For very thick and compactly fluorescent samples, however, confocal microscopy has an advantage in that its pinhole removes out-of focus light physically. Structured illumination microscopy is quite effective at removing out-of-focus light computationally, because it is not subject to the missing-cone problem, but computational removal leaves behind the associated shot noise. Therefore confocal microscopy may be preferable on very thick and dense samples, for which the in-focus information in a conventional microscope image would be overwhelmed by out-of-focus light, whereas structured illumination microscopy may be superior in a regime of thinner or sparser samples.' This point is not mentioned at all in the manuscript, yet we are certain it is at least partially responsible for the residual image artifacts the authors mention. Please discuss the problem of out of focus light on 3D samples, particularly with an eye to the 'spot-scanning' papers mentioned above.

    Response:

    We appreciate this significant obstacle and we want to thank Reviewer #1 for emphasising its importance. To address the comment, we plan to add a discussion of the significance of out-of-focus light to SIM imaging to the introduction, results, and discussion sections of the manuscript.

    1. The authors use a water dipping lens, yet they image into samples that are mounted on coverslips, i.e. they use a dipping lens to image through a coverslip:

    This almost certainly introduces spherical aberration, which the authors seem to observe: see attached pdf for reference

    We find this troubling, as it seems that in the process of building their setup, the authors have made a choice of objective lens that introduces aberrations - that they later correct. At the very least, this point needs to be acknowledged in the manuscript (or please correct us if we're wrong) - as it renders the data in Figs. 3-4 somewhat less compelling than if the authors used an objective lens that allowed correction through a coverglass, e.g. a water dipping lens with a correction collar. In other words, in the process of building their AO setup, the authors have introduced system aberrations that render the comparison with 3D SIM somewhat unfair. Ideally the authors would show a comparison with an objective lens that can image through a glass coverslip.

    Response:

    We want to thank Reviewer #1 for raising this point, which we did not describe clearly enough, leading to confusion. We should have made it clearer that we used a water dipping/immersion objective lens with a correction collar which extends from no coverslip (dipping) up to well beyond a standard #1.5 (170 um thick) coverslip. We adjusted this collar before each image acquisition session, to ensure that the system is optimised for each experiment individually and that the spherical aberrations are minimal before any DM-based correction. We plan to elaborate and emphasise this point in several places in the revised manuscript, including in the figure legends, materials and methods and results sections, to avoid any ambiguity and confusion about the use of the correction collar and this particular water immersion/dipping objective lens.

    1. The authors tend to include numbers for resolution without statistics. This renders the comparisons meaningless in my opinion; ideally every number would have a mean and error bar associated with it. We have included specific examples in the minor comments below.

    Response:

    This is a good point, which we address below, in three minor comments. In summary, to address this comment, we plan to include statistical information in the revised manuscript.

    1. In Fig. 5, after the 'multipoint AO SIM', the SNR in some regions seems to decrease after AO: see attached pdf for reference

    Please comment on this issue.

    Response:

    We want to thank Reviewer #1 for the insightful comment. There are multiple phenomena in effect here, which cause the drop in intensity. The most prominent one is photobleaching, as the AO image stack (right) was acquired after the bypass one (left). To address this comment, we plan to add additional data and to include a brief discussion about this issue and other related points.

    1. Please provide timing costs for the indirect AO methods used in the paper, so the reader understands how this time compares to the time required for taking a 3D SIM stack. In a similar vein, the authors in Lines 213-215, mention a 'disproportionate measurement time' when referring to the time required for AO correction at each plane - providing numbers here would be very useful to a reader, so they can judge for themselves what this means. What is the measurement time, why is it so long, and how does it compare to the time for 3D SIM? It would also be useful to provide a comparison between the time needed for AO correction at each (or two) planes without remote focusing (RF) vs. with RF, so the reader understands the relative temporal contributions of each part of the method. We would suggest, for the data shown in Fig. 5, to report a) the time to acquire the whole stack without AO (3D SIM only); b) the time to acquire the data as shown; c) the time to acquire the AO stack without RF. This would help bolster the case for remote focusing in general; as is we are not sure we buy that this is a capability worth having, at least for the data shown in this paper.

    Response:

    We agree that the timing (and other) costs can be an important consideration, and we want to thank Reviewer #1 for bringing up this good point. To address this issue, we plan to expand our description of the AO methods, also including numbers for the time it takes to perform the different parts. In terms of comparisons, the RF makes no contribution to the timing costs of the aberration correction, a point that we want to make clearer in the results and the methods and materials sections, as the two are independent processes in our approach. Instead, the RF can be compared to standard focusing with a piezo stage, a point which we discuss in the supplementary material. We plan to make this point clearer in the discussion section of the main manuscript, and to emphasise better the advantages of the RF in terms of imaging speed.

    1. Some further discussion on possibly extending the remote focusing range would be helpful. We gather that limitations arose from an older model of the DM being used, due to creep effects. We also gather from the SI that edge effects at the periphery of the DM was also problematic. Are these limitations likely non-issues with modern DMs, and how much range could one reasonably expect to achieve as a result? We are wondering if the 10 um range is a fundamental practical limitation or if in principle it could be extended with commercial DMs.

    Response:

    Regrettably, we were not able to try other DMs on the Deep3DSIM system. However, Jiahe and colleagues show in [1] that similar DM-based remote focusing, even with the same model deformable mirror, can be pushed to 120 um (Strehl ratio >0.8) with a 0.42 NA dry lens (20 mm WD) and close-loop wavefront compensation operation. While this is not directly translatable to high NA 3D-SIM imaging, we expect that with a stable version of the same DM the useable RF range could be easily increased twice or even more. We thank Reviewer #1 for the good comment, which we plan to address by revising the text to make the limitations clearer and by citing relevant studies.

    [1] Cui, J., Turcotte, R., Emptage, N. J., & Booth, M. J. (2021). Extended range and aberration-free autofocusing via remote focusing and sequence-dependent learning. Optics Express, 29(22), 36660-36674.

    Minor comments:

    1. The paper mentions Ephys multiple times, even putting micromanipulators into Fig. 1 - although it is not actually used in this paper. If including in Figure 1, please make it clear that these additional components are aspirational and not actually used in the paper.

    Response:

    Although not shown in the context of this paper, the Deep3DSIM system was built specifically around experiments such as electrophysiology, which can benefit from the upright configuration and the water-dipping-capable objective lens. To address this comment, we plan to clarify the role of the micromanipulators and to update Figure 1 accordingly.

    1. The abstract mentions '3D SIM microscopes', 'microscopes' redundant as the 'm' in 'SIM' stands for 'microscope'.

    Response:

    We accept that “3D SIM microscopes” sounds repetitious and we plan to revise the wording of the abstract to “3D SIM system”.

    1. 'fast optical sectioning', line 42, how can optical sectioning be 'fast'? Do they mean rapid imaging with optical sectinong?

    Response:

    Yes, we meant rapid imaging with optical sectioning. We plan to change the wording to make it less ambiguous.

    1. line 59, 'effective imaging depth may be increased to some extent using silicone immersion objectives', what about water immersion objectives? We would guess these could also be used.

    Response:

    Yes, water immersion objective lenses also fall in the same category and we plan to rephrase this part to state it explicitly.

    1. line 65 - evidence for 'water-dipping objectives are more sensitive to aberrations' ? Please provide citation or remove. They are certainly more prone to aberrations if used with a coverslip as done here.

    Response:

    The refractive index (RI) of cells and tissues [1] is closer to the RI of silicone oil (~1.4) than it is to water (~1.33). Therefore, because of the larger difference in RI, imaging with a water-dipping objective lens is more prone to aberrations from RI mismatch. We plan to rephrase this argument to make it clearer.

    [1] Jacques, S. L. (2013). Optical properties of biological tissues: a review. Physics in Medicine & Biology, 58(11), R37.

    1. 'fast z stacks' is mentioned in line 103. How fast is fast?

    Response:

    The speed would depend on the way Z-stacks are being acquired. For example, acquisitions with two channels would be at least twice as fast, because of the ability to do simultaneous imaging on the Deep3DSIM system. Likewise, experiments that can benefit from the remote focusing can be several times faster than using a Z piezo stage, and this point is discussed in the supplementary material (section “Step response”). Finally, thanks to the electronic design of the imaging system, orchestrating everything via digital logic (e.g. TTL) signals, and thanks to the elaborate control software, we can ensure that all image acquisitions are carried out as quickly as possible, operating near the limit of the underlying hardware devices. We plan to explain these points in a clear way in the discussion section, and we plan to provide more numbers in the supplementary material.

    1. line 116 'we imaged 100 nm diameter green fluorescent beads'. Deposited on glass? Given that this paper is about imaging deep this detail seems worth specifying in the main text.

    Response:

    Yes, in this case the beads were deposited on glass. We plan to include this detail in the description of the experiment.

    1. lines 127-130, when describing changes in the bead shape with numbers for the FWHM, please provide statistics - quoting single numbers for comparison is almost useless and we cannot conclude that there is a meaningful improvement without statistics.

    Response:

    We agree with this comment. We plan to include statistical information for all the FWHM numbers.

    1. In the same vein, how can we understand that remote focus actually improves the axial FWHM of the widefield bead? Is this result repeatable, or it just noise?

    Response:

    The lower axial FWHM with remote focusing is likely caused by data fitting or quantification error. Together with the inclusion of statistical information, we plan to review all the resolution values and to ensure that they are accurate and sensible.

    1. line 155, 'Because of the high spatial information...' -> 'Because of the high resolution spatial information...'

    Response:

    We agree with this comment. To address it, we plan to rephase this part.

    1. When quoting estimated resolution #s from microtubules (lines 158-163) similarly please provide statistics as for beads.

    Response:

    We agree with this comment. To address it, we plan to include statistical information for the resolution values from microtubules.

    1. It seems worth mentioning the mechanism of AO correction (i.e. indirect sensing) in the main body of the text, not just the methods.

    Response:

    We agree with this comment. To address it, we plan to describe briefly the aberration correction method in the introduction or the results section.

    1. How long do the AO corrections take for the datasets in the paper?

    Response:

    The duration of the aberration correction routines is directly proportional to the number of Zernike modes, the number of iterations, the exposure time of the camera, and other parameters. In our experiments, it was usually in the order of tens of seconds. To address this comment, and in line with the sixth major comment, we plan to include more details about the timing of the different parts of the AO methods.

    1. Were the datasets in Fig. 2-4 acquired with remote focusing, or in conventional z stack mode? Please clarify this point in the main text and the figure captions.

    Response:

    The only data acquired with RF in Fig. 2-4 are one bead in Fig. 2A and another bead in Fig. 2B, both labelled accordingly. We plan to make it clearer in the text that the rest of Figure 2, as well as Figures 3 and 4, were acquired with the piezo Z stage.

    1. It would be helpful when showing z projections in Figs. 3-5 to indicate the direction of increasing depth (we assume this is 'down' due to the upright setup, but this would be good to clarify)

    Response:

    The direction is indicated by the arrows labelled with ‘Z’. We plan to clarify this in the figure captions.

    1. line 174, 'showed significant improvements in both intensity and contrast after reconstruction' - we see the improvements in contrast and resolution, it is harder to appreciate improvements in intensity. Perhaps if the authors showed some line profiles or otherwise quantified intensity this would be easier to appreciate.

    Response:

    We agree with this comment. To address it, we plan to change Figure 3 to illustrate the improvement in intensity, likely with line profiles, as suggested by the reviewer.

    1. line 195 'reduced artefacts' due to AO. We would agree with this statement - the benefit from AO is obvious, and yet there are still artefacts. If the authors could clarify what these (residual) artefacts are, and their cause (out of focus light, uncorrected residual aberrations, etc) this would be helpful for a reader that is not used to looking at 3D SIM images.

    Response:

    We agree with this comment. To address it, we plan to explain this point in both the results and the discussion sections.

    1. Line 197, 'expected overall structure', please clarify what is expected about the structure and why.

    Response:

    We agree with this comment. To address it, we plan to describe better the Canoe (Cno) protein, including an explanation of its expression pattern, which is the honeycomb-like structure observed in the images.

    1. Line 199, what is a 'pseudo structure'?

    Response:

    We used this expression to refer to unclear (e.g. dim, fuzzy) structures. We plan to improve the wording of that part of the results section.

    1. Fig. 4B, 'a resolution of ~200 nm is retained at depth', please clarify how this estimate was obtained, ideally with statistics.

    Response:

    We agree with this comment. To address it, we plan to clarify this point in the results section, including statistical information.

    1. Fig. 4D, please comment on the unphysical negative valued intensities in Fig. 4D, ideally explaining their presence in the caption. It would also be helpful to highlight where in the figure these plots arise, so the reader can visually follow along.

    Response:

    We agree with this comment. To address it, we plan to explain how negative intensities arise in SIM reconstruction, often a result of spherical aberrations, and we plan to indicate where the line profile in Figure 4D comes from.

    1. Line 245, 'rapid mitosis'. What does rapid mean, i.e. please provide the expected timescale for mitosis.

    Response:

    The mitotic cycles at this developmental stage are short, e.g. 5 minutes per mitosis, compared to those of somatic cells where it takes several hours. We plan to include this information in the main text.

    1. For the data in Fig. 6, was remote refocusing necessary?

    Response:

    Yes, it was necessary because the point of Figure 6 is to demonstrate the combination of remote focusing and SIM super-resolution in live samples. Drosophila embryos are a very good sample for this kind of demonstration, because they are often subject to micromanipulation (e.g. injection and electrophysiology), and these are the kind of experiments that can benefit greatly from the optical axial scanning of the remote focusing, where the sample can remain stationary. However, there is nothing preventing the imaging of this kind of sample with a piezo Z stage or with some other kind of mechanical actuator. In this sense, the remote focusing is not strictly necessary but still much more convenient in some applications. We plan to make this point clearer in the discussion section.

    1. What is the evidence for 'reduced residual aberrations', was a comparative stack taken without AO? In general we feel that the results shown in Fig. 6 would be stronger if there were comparative results shown without AO (or remote focusing).

    Response:

    We agree with this comment. In general, it is difficult to make direct comparisons (e.g. as in Figures 3-5) with live samples, because of the dynamic character of the samples, where it is often impossible to capture the same scene more than once. To address this comment, we plan to revise the wording of the relevant part of the results section, to ensure that the data in Figure 6 is properly described.

    1. Line 350, 'incorporation of denoising algorithms' - citations would be helpful here.

    Response:

    We agree with this comment. To address it, we plan to add references to the relevant statement, showing examples of denoising in 3D-SIM imaging and reconstruction.

    1. Line 411, 'All three were further developed and improved' - vague, how so?

    Response:

    A detailed breakdown of all the changes is available on the respective software repositories. We also plan to add a summary in the supplementary material.

    1. Sensorless AO description; how many Zernike modes were corrected?

    Response:

    We usually corrected 8 modes: Z5 to Z11 and Z22, using Noll indexing. We plan to add a table to the supplementary material, describing which modes were corrected for each dataset.

    1. Multi-position aberration correction. Was the assumption of linearity in the Zernike correction verified or met? Why is this a reasonable assumption?

    Response:

    By their very definition, some aberrations, such as defocus and spherical aberrations, change linearly with depth. Others are also proportional to the imaging depth, and first-order approximation (i.e. straight line) is the most sensible for just two correction points, as is the case with the dataset presented in Figure 5. We plan to explain this point better in the results section.

    1. Fig. S1B is not useful; if the idea is to give a visual impression of the setup, we would recommend providing more photos with approximate distances indicated so that the reader has a sense of the scale of the setup. As is - it looks like a photograph of some generic optical setup.

    Response:

    We agree with this comment. To address it, we plan on including more photos in the supplementary material, to give a better sense of the scale.

    1. SI pattern generation - 'the maximum achievable reconstruction resolution was only slightly reduced to about 95% of the theoretical maximum'. We don't understand this sentence, as the resolution obtained on the 100 nm beads is considerably worse than 95% of the theoretical maximum. Or do the authors mean 95% of the theoretical maximum given their pitch size of 317 nm for green and 367 nm for red?

    Response:

    Limiting the stripe width to about 90% of what is achievable leads to a reduction of the theoretical maximum resolution to 95% of what it could be. We plan to rephrase this part to make it clearer.

    1. SI Deformable mirror calibration 'spanning the range [0.1, 0.9]' - what are the units here?

    Response:

    These are normalised control amplitudes, i.e. [10%, 90%], which means that they are unitless. We plan to explain this in a clearer way.

    1. What are the units in Fig. S5C, S5D?

    Response:

    Errors are in radians, defined by the calibration interferometric wavefront sensor. We plan on updating the figure to include this information.

    1. It would be useful to define 'warmup' also in the caption of SI Fig. S6A.

    Response:

    We agree with this comment. We plan to change the caption of Figure S6A to clarify this point.

    1. SI Remote Focusing, 'four offsets, {-5 mm, -2.5 mm, 2.5 mm, 5 mm}...' are the units mm or um?

    Response:

    The units are supposed to be um (micrometres). We plan on fixing this error.

    1. '...whereas that of the 10 beads was...' here, do the authors mean the position of the beads derived from the movement of the piezo stage, as opposed to the remote focusing?

    Response:

    This is the average standard deviation between the 10 different beads, all from volumes acquired with remote focusing. We plan on rephrasing this part to make it clearer.

    1. The authors refer to the 'results from Chapter 3.2'. What are they talking about? Do they mean a supplementary figure, or earlier supplementary results? In general, we found the discussion in this paragraph difficult to follow.

    Response:

    This is a remnant from an earlier version of the document which used numbered sectioning. Chapter 3.2 is referring to the section titled “Characterisation of drift and temperature effects”. We plan on revising this paragraph to make it clearer.

    1. Supplementary Fig. 9 seems to be not referred to anywhere in the text.

    Response:

    We agree with this comment. To address this issue, we plan on referring to this figure in the main text.

    1. Since the paper emphasizes 3D SIM, OTFs along the axial direction would also be useful to show, in addition to the lateral OTFs shown in Fig. 2D.

    Response:

    We agree with this comment. To address it, we plan on adding orthogonal views of the OTFs to the supplementary material.

    1. When the sample is moved by the piezo, the axial phase of the 3D-SIM illumination pattern is stable as the sample is scanned through the illumination pattern. When remote focusing is performed, the sample is always stable so the axial phase of the 3D-SIM illumination pattern is presumably changing with remote focusing. Can the authors clarify if the 3D SIM illumination pattern is scanned when remote focusing is applied, or is the intensity pattern stable in z?

    Response:

    Yes, the illumination pattern is scanned. We plan on clarifying how the structured illumination works in the case of remote focusing in the supplementary material.

    1. In Supplementary Fig. 9, primary spherical is referred to twice, both at index 11 and 22. The latter is presumably secondary spherical?

    Response:

    Yes, it is supposed to be secondary spherical aberrations. We plan on fixing this error.

    1. we do not understand the x axis label, in Fig. S4D, is it really [0, 50, 50, 50] as written?

    Response:

    The labels of the x-axis are not well formatted. There are three range of [0, 50] where only the first zero is properly displayed. We will revise this part of the figure to make it clear.

    Reviewer #2

    1. The authors have provided an incomplete description of the structured illumination microscopy (SIM) reconstruction process. It is unclear whether the approach is based on 2D interference SIM configurations or 3D interference patterns. Furthermore, the specific algorithm utilized for image reconstruction has not been elucidated. Elaborating on these aspects is crucial as they significantly influence the interpretation of the resulting data.

    Response:

    We want to thank Reviewer #2 for bringing our attention to the incomplete description of the reconstruction process. Our approach was based on 3D interference patterns and it was carried out using the Gustafsson’s reconstruction techniques as implemented by the softWoRx software, designed for the OMX 3D-SIM microscopes. To address this comment, we plan to revise the manuscript and to include more details about the 3D-SIM reconstruction techniques in the methods and materials section.

    1. The authors have stated that sample-induced aberrations caused by RI inhomogeneities within the specimen is another major reason for causing artifacts generation. Literature has demonstrated that RI inhomogeneities can lead to non-local distortions in the grid pattern, which suggests that applying uniform reconstruction parameters across the entire image may not be viable. Traditional artifact remediation using the classical Wiener method is likely insufficient under these conditions (PMID: 33896197). The existing adaptive optics (AO) approach, which employs a deformable mirror (DM) alongside an sCMOS camera, is inadequate for tackling the issue at hand. Actually the assertion made in the paper that "aberrations change approximately linearly with depth" is seemingly contradicted by simulations referenced in the cited literature (PMID: 33896197). Consequently, it appears that the current methodology might only achieve a partial mitigation of the problems associated with spherical aberration resulting from RI mismatches. It is advisable, therefore, that the authors explicitly acknowledge this limitation in their manuscript to prevent any potential misinterpretation by readers.

    Response:

    We are thankful for the thoughtful comment by Reviewer #2. The focus of our work was not the use of advanced 3D-SIM reconstruction and aberration correction methods; instead, we used standard ones which are not able to deal perfectly with anisoplanitism, i.e. when the aberrations vary laterally. As such, our approach provides an average reconstruction and correction across the field of view. In our particular setup this anisoplanitism was not very significant, but we agree that it could be an issue for optical systems with very wide field of view. To address this good point, we plan on clarifying these potential issues in the results and the discussion sections.

    1. In Figure 2, the use of COS-7 cells, which are known for their relatively thin axial dimension, for the experiments raises an eyebrow. Notably, there are ample instances in existing research where both 2D-SIM and 3D-SIM, without the integration of adaptive optics, have yielded high-quality super-resolution images of structures such as tubulin and the endoplasmic reticulum. In addition, the authors did not present a direct comparison between BP-SIM and AO-SIM here. Without this comparative analysis, it remains ambiguous whether the enhancements in resolution and contrast and the reduction in artifacts can genuinely be attributed to the mitigation of spherical aberration. To clarify this, it would be beneficial for the authors to include side-by-side comparisons of these modalities to demonstrate the specific improvements attributed to AO-SIM.

    Response:

    We are grateful to Reviewer #2 for this helpful comment. In Figure 2, we demonstrate the performance we get out of 3D-SIM in terms of optical resolution. We do not make any statements about the impact of the aberration correction on image quality. Nevertheless, to address this comment, we plan to revise the figure to explain more clearly and explicitly this point.

    1. In Figures 3 and 4, the authors have illustrated the enhancements achieved through the application of AO. However, there is a discernible presence of hammer-stroke and honeycomb artifacts in pre-AO imaged data, which seem to originate from the amplification of the incorrectly moved out-of-focal background in the frequency domain. Various strategies have been previously suggested to address these specific artifacts, encompassing methods like subtracting background noise in the raw images or employing selective frequency spectrum attenuation techniques, such as Notch filtering and High-Fidelity SIM. To facilitate a more comprehensive understanding, I would recommend that the authors incorporate into their study a comparison that includes BP-SIM data that has undergone either background subtraction or frequency spectrum attenuation. This added data would enable a more complete evaluation and comparison regarding the merits and impact of their AO approach.

    Response:

    We thank the reviewer for this excellent suggestion and we agree that a pre-processing step, such as background subtraction or frequency spectrum attenuation, can help with the reduction of artefacts. To address this comment, we will re-analyse our data and apply these techniques, and we will add the data to the manuscript, with an appropriate revision to the text.

    Reviewer #3

    1. There is an overall reference in the manuscript of the novelty possible range of applications of using an upright microscope configuration. Examples mentioned are tissue-based imaging, access to whole-mount specimens for manipulation and electrophysiology. However, authors fail to present any such applications. There is not a single example presented which could not have been obtained with an inverted microscope. Could the authors provide an example where a water-dipping is used. Expanded samples could be one case, since the thickness of the gel makes it difficult to image with an inverted microscope. Another possible example would be to label the extracellular space and do shadow imaging of the tissue (SUSHI PMID: 29474910). ExM might be simpler to do as part of revising the manuscript than SUSHI.

    Response:

    We are thankful to Reviewer #3 for these interesting comments. To address this comment, we will emphasise more clearly that Figure 6 of our manuscript shows a sample that is often part of live imaging experiments that require microinjection and even electrophysiology. Our aim was to show the proof of principle and the potential of such experiments, rather than to carry out real and complex experiments using electrophysiology or microinjection. Regarding providing an example where water-dipping is used, this is already present in the same Figure 6, which we will describe more explicitly and fully in the revised manuscript. The reviewer’s comments on expansion microscopy and SUSHI are interesting, but the primary purpose of our microscope system is to facilitate super resolution live cell imaging experiments. Nevertheless, to address this comment, we will add an explanation of the relevance of our approach to improving deep super resolution imaging of expanded specimens.

    1. On the main text it is described a 5-fold volumetric resolution, which is confusing since authors only mention lateral and axial resolutions. Their measurements correspond to a ~1.6-fold lateral improvement and ~1.7-fold axial improvement. These are however not the 95% of the achievable resolution theoretical maximum, as stated in p7 SI (2 fold increase of 282nm), but only the 80-85%. This point should be rephrased in the manuscript.

    Response:

    We want to thank Reviewer #3 for bringing up this important point. To address it, we plan to make changes to the text, both in the main manuscript and in the supplementary material, to make it clearer what are the resolution improvement that we achieve and what are the limitations to our approach.

    1. [OPTIONAL] p4 and related to figure 2, it would be important to report also measurements of beads with SIM but without AO, just as done for WF. Is there an improvement of using AO on SIM? This is reported for the fixed cells but not for the beads.

    Response:

    We found no significant improvement in resolution when AO was applied to SIM. To address this comment, we plan to add the extra data to Figure 2, demonstrating this point.

    1. Figure 2, it is odd the comparison between WF+/- AO and SIM +/- AO are done using different cellular structures. Since wavelengths used are not the same it is difficult to interpret if there is any improvement of using AO on SIM compared to SIM without AO. Same questions arise as above, Is there an improvement of using AO on SIM?

    Response:

    We agree that the data in Figure 2C and 2D is presented in unusual way. Our intention was not to make a comparison between bypass and AO, but instead to characterise the super-resolution capabilities of the system. We use different channels because doing -/+ AO consecutively leads to noticeable intensity drop due to photobleaching. We are grateful to Reviewer #3 for the valuable comment, which we plan to address by revising Figure 2.

    1. "A significant benefit and uniqueness of the Deep3DSIM design is its upright configuration, whereas commercial SIM systems are built around inverted microscopes and are usually restricted to imaging thin samples, such as cultured cells." (p5) is not correct. The commercial DeepSIM module from CREST Optics can be mounted on an inverted microscope as well as image deep into tissue (seehttps://crestoptics.com/deepsim/ and application notes therein) and be used with essentially any objective. This point should be rephrased in the text.

    Response:

    We want to thank Reviewer #3 for bringing our attention to this error. Of course, we meant commercial 3D-SIM systems, such as GE Healthcare DeltaVision OMX and Nikon N-SIM. To address this issue, we plan to rephrase this part of the results section. Regarding the commercial DeepSIM module from CREST Optics, as far as we can tell, it uses a different method – 2D lattice multi-spot SIM – which comes at the cost of signal loss when sample-induced aberrations are strong. This is very different from our method, which uses a deformable mirror to manipulate the phase information of both the excitation and the emission light at the back-pupil plane of the objective lens, which can theoretically provide 2× resolution enhancement with no signal lost.

    1. Fig 3 reports the improvements of AO on SIM for imaging over 10um in tissue. What are the Zernike modes measured? Or how does the pupil look like before and after correction? It would be also good to report the Fourier amplitudes as done in Fig 2C as a quantitative measure of improvement. It would be good to point out the artifacts observed on the BP SIM image reconstruction (labelled with 3x, fringes are noticeable).

    Response:

    We thank Reviewer #3 for the good suggestions. We plan to add information about the measured Zernike modes to the results section, as well as to add a brief discussion about the noticeable reconstruction artefacts. In terms of pupil and Fourier amplitudes, we plan to change Figure 3 to include all this information or, alternatively, to include it in the supplementary material.

    1. Many key details relating to image acquisition and AO correction are missing for all figures. How is the AO optimization implemented? Is it implemented via a genetic algorithm (progressive optimization of parameters) or using more clever strategies? Not clear if the optimization is implemented using images obtained with flat illumination or after SIM imaging/processing of a given dataset. How long does the AO optimization take? How sensitive to noise is the process? What metric do they use to estimate the sensorless AO correction? On pag12, they say "Fourier domain image metric" for measurements with fine details; otherwise, ISOsense when not high frequencies are present. Could the authors report the formula used to calculate the first metric? What do they consider to be low and high frequencies in this case? Is there a reason why ISOsense is not always used, or is there an automatic way to choose between the two? How many images were acquired for AO correction? Which samples were corrected with ISOsense and which ones with Fourier domain image metric? (see for example the detailed experimental reporting in the Supp Mat from Lin et al Nat Commun 2021).

    Response:

    We are grateful to Reviewer #3 for the extensive list of questions. The optimisation is done via non-linear least square, it uses widefield images, and it is performed before the actual image acquisition, i.e. well before any SIM reconstruction takes place. The methods used for aberration correction are described in the Methods and materials section, and further in the cited literature, e.g. Antonello et al 2020 and Hall et al 2020. ISOsense needs to be manually chosen over the Fourier image metric, and this should be done when large mode biases lead to small changes in the metric value, which is likely to happen when there are little or no sharp features in the images. One of the disadvantages of our implementation of ISOsense is that the structured illumination pattern is continuously exposed over the sample, which leads to photobleaching and phototoxicity. None of the datasets shown in the manuscript use ISOsense. To address all of the questions from this comment, we plan to significantly expand our descriptions of the AO methods, both in the main text and in the supplementary material.

    1. Fig 4. Data presented for larval brain tissue is a very clear example of adding AO to image deep into tissue as the effect at ~130 cannot be understated. Here too, it would be also good to report the Fourier amplitudes as done in Fig 2C as a quantitative measure of improvement and possibly the SNR of reconstructed images. Having a way to quantitatively describe how much better are those images would be great. Also, what are the aberrations corrected? Can the wavefront or Zernike amplitude of the modes be reported? Same as for Fig 3, details about AO correction are missing.

    Response:

    We are grateful to Reviewer #3 for the helpful comment. We will address it by adding the Fourier amplitudes to Figure 4, as suggested, and by reporting the Zernike mode amplitudes of the aberration corrections.

    1. [OPTIONAL] "It is worth noting that aberrations can differ across larger fields, and therefore, after applying an average correction, residual aberrations can still be observed in some regions of the AO-corrected bead images. However, the overall PSF shapes were still dramatically improved with AO compared to the equivalent without AO." This point is very interesting although not result either in the main text or in the SI is presented.

    Response:

    The residual aberrations are present in the right image of Figure 4B, although we did not highlight them specifically. We are thankful to Reviewer #3 for the good suggestion and we plan to implement it by changing Figure 4 to show a few of the beads with residual aberrations.

    1. "As we found that the aberrations change approximately linearly in depth, we could measure the aberration in two planes and calculate the corrections in intermediate planes by interpolation, an approach which we termed "multi-position AO"." This is, personally, one of the major contributions of this work to the community. Unfortunately, it is not reported in detail. Not only for SIM but for imaging with WF or confocal, such linear change for aberrations with depth is not well known. Again, here the details of AO correction and image metrics are missing. To establish that for most thick biological structures 'aberrations change approximately linearly in depth' would be foundational to the widespread use of AO within standard imaging. Would it be possible for the authors to elaborate on this point and present detailed results? What is the error from measuring and correcting more than 2 planes? What is the error from just measuring and AO correcting at the deeper plane, i.e. from a single measurement? Authors could also show a case in which a linear assumption works nicely (or how well it works). For example, comparing an intermediate plane (or a plane beyond) imaged after AO optimization or after coefficient interpolation of the Zernike modes and compare it against correcting directly that plane.

    Response:

    Some aberrations, such as defocus and spherical aberrations, are mathematically defined as varying linearly with depth. The change in other aberrations with depth can also be estimated with a linear model, which is a standard first-order approximation in the case of two datapoints, such as corrections done in Figure 5. It is not possible to do regression analysis with just a single point, so it is impossible to apply our multi-position AO at a single plane. We are grateful to Reviewer #3 for the constructive comment. To address the questions in this comment, we plan to provide a more detailed description of the correction estimation methods to the results section, as well as a discussion on the accuracy of the linear model in the discussion section.

    1. The image of the cos-7 cell in metaphase, for Fig 5 is, however, very disappointing. See Fig 1 of Novak et al Nat Commun 2018 for an example of a single z-plane of a cell in metaphase. Having the possibility to correct for the entire 3D volume, I would expect amazing 3D volumes (movies and/or projections) associated with this imaging which are not presented.

    Response:

    We thank Reviewer #3 for the interesting comment. The example in Novak et al 2018 was acquired with STED microscopy, which is an entirely different imaging method and thus produces different results. Nevertheless, we will revise the discussion of Figure 5 to ensure that the right expectations are set.

    1. In Figure 6, they use AO in remote configuration mode to allow imaging of live specimens. It needs to be clarified if this is an a priori characterization that is then kept fixed while recording in time. The last acquired volume of fig 6A and B have a higher amount of artifacts with respect to time 00:00. Are those artifacts due to lower SNR (maybe due to sample bleaching) or due to some change in the aberrations of the specimen?

    Response:

    We want to thank Reviewer #3 for the valuable comment. We assume that by change in artefacts, Reviewer #3 is referring to the overall green fluorescent structure. Indeed, this last volume shows the anaphase to telophase transition where the mitotic spindle is being reorganised and disassembled. As such, the structure is much less well-defined than in the first volume. The changes in aberrations over time are not particularly significant in this case, and the photobleaching is not that impactful in such an experiment where relatively thin volumes are acquired with substantial time delay between them. To address this comment, we plan to revise the discussion of the figure and to ensure that the scene observed in the last volume is clearer.

    1. "These results demonstrate that the remote focusing functionality of the system can be successfully applied for live 3D-SIM experiments, allowing four-dimensional acquisition while keeping the specimen stationary, thus avoiding the usual agitation and perturbations associated with mechanical actuation." Generally, this statement is true, but for the specific example shown of drosophila embryogenesis is it relevant? If they use piezo-driven Z-stack imaging with AO, does that lead to incorrect reconstructions or motion-induced artifacts? Related to the results shown in Fig 6, the fair comparison would be AO SIM vs SIM (without AO), not AO SIM vs AO WF.

    Response:

    We are grateful to Reviewer #3 for the insightful comment. Drosophila embryos are quite robust to perturbations due to their shape and size, and the restrictions imposed by SIM experiments (e.g. small Z steps and Z levels held for long periods of time) make motion-induced artefacts not very impactful. Regarding the results, the point of Figure 6 is not to demonstrate the advantages of aberration correction, which we do not claim in the caption or in the relevant part of the discussion, but to demonstrate that remote focusing works well with 3D-SIM reconstruction, which is known to have stringent requirements about the image quality. To address this comment, we plan to revise the figure and its relevant part of the results section.

    1. When performing remote focusing, is the effective NA of the imaged plane changing with respect to the NA of the objective used at its focal plane?

    Response:

    We thank Reviewer #3 for the good question. The effective NA is not altered by the remote focusing. We plan to mention this detail in the results section.

    1. [OPTIONAL] Did the authors run calculations to explore whether a commercial upright microscope could be used instead of their design? Are there any fundamental flaws that would make impossible using a commercial base? If not, could an AO SIM module be designed such that it adds on a commercial base? It would be important to discuss this point.

    Response:

    We thank Reviewer #3 for bringing up this interesting point. A lot of considerations, calculations, and modelling were done in the design of the Deep3DSIM system. Of course, the use of a commercial upright microscope stand was part of the deliberation. One of the obvious limitations is the difficult access to the pupil-conjugated plane. On the other hand, a commercial microscope stand is not well compatible with many of the key parts of the system, which were designed around specific biological applications, such as dual camera system for fast live simultaneous imaging and the heavy-duty Z stage intended to support two heavy micromanipulators. To address this comment, we plan to add a discussion of the compatibility of Deep3DSIM with commercial microscope stands to the discussion section and the supplementary material.

    Minor comments:

    1. Fig 2 lacks a color bar for D panels, which is in log scale. Authors should also show the Fourier transform along the z direction.

    Response:

    The colour mapping in Figure 2 uses the lookup tables called Cyan Hot and Orange Hot, as indicated in the caption, which come from the ImageJ software. To address this comment, we plan to improve the caption to reflect the fact that the plots are in log scale. We also want to include Fourier transforms along Z, either in the figure itself or in the supplementary material.

    1. p4, "Such minor aberrations tend to be insignificant in conventional microscopy modalities such as widefield and confocal (Wang and Zhang, 2021). Therefore..." If optical aberrations are insignificant for single cells in widefield and confocal why do experiments here? These sentences should be rephrased to motivate better the experiments performed.

    Response:

    We agree with this comment. To address it, we plan to rephrase this part of the results section to motivate better the experiments.

    1. Imaged microtubules look abnormal, 'dotty' (figure 2) in both WF and SIM. See https://sim.hms.harvard.edu/portfolio/microtubules/ or Fig 1 of Wegel, et al Dobbie Sci Rep 2016, for better examples of continuous microtubule structures as imaged with SIM.

    Response:

    The dottiness of the microtubule structures is not related to the SIM reconstruction, because the same dottiness is seen in the respective WF data, too. It is a product of the sample preparation and it has only aesthetic significance. Nevertheless, to address this comment we plan to mention the dottiness in the results section.

    1. Is also the remote focusing performed via optimization of metrics similar to the one used for compensating aberrations?

    Response:

    Yes, as mentioned in the Methods and materials (p. 13), the calibration of the remote focusing involved sensorless aberration correction of several Zernike modes, such as defocus and spherical aberrations.

    1. Figure 2, the order of names on the top right of the panel should match the order of curves presented.

    Response:

    We agree with this comment. To address it, we plan to reorder the curves in Figure 2.

    1. I value the efforts to improve open-source tools for system and AO control and GUI. And those tools seemed to have been modified for this work, although those modifications are not described. Would it be possible for the authors to describe those modifications?

    Response:

    A detailed breakdown is publicly available at the respective software repositories. To address this comment, we plan to add a summary of software changes to the supplementary material.

    1. Reported average values of the FWHM of imaged beads in 3D (p4) require also to report errors associated with those measurements.

    Response:

    We agree with this comment. To address it, we plan to add statistical information to the FWHM values on page 4.

    1. Page 13, second paragraph states that "The results from chapter 3.2..." I believe that was a copy/paste from a thesis but should be corrected for a peer-reviewed publication, as there is no chapter 3.2.

    Response:

    This is a leftover from an older version of the document which used numbered sectioning. In this case “chapter 3.2” refers to subsection “Characterisation of drift and temperature effects”. We plan on fixing this mistake in the revised manuscript.

  2. Review Commons

    Note: This preprint has been reviewed by subject experts for Review Commons. Content has not been altered except for formatting.

    Learn more at Review Commons

    Referee #3

    Evidence, reproducibility and clarity

    Summary

    The work of Wang et al entitled "Deep super-resolution imaging of thick tissue using structured illumination with adaptive optics" presents the use of a deformable mirror to simultaneously perform adaptive optics 'AO' and remote focusing 'RF' on a custom-designed upright microscope configuration. The work is novel and represents a timely application of this type of technology to imaging biological specimens. AO enables the correction of refractive index mismatch and sample-induced aberrations while remote focusing allows focusing through the sample without moving the specimen or the objective. The use of AO improves the final image reconstructed using a traditional SIM processing strategy. I greatly value the idea presented (SI pattern generation p7 Supp Inf) about maximizing the contrast of the projected structured illumination. This could be an excellent way to improve SIM imaging since image reconstructions suffer from artifacts when signal to noise ratio is low. However, since it is only one of the factors considered for reducing the stripe width it is unclear how it compares to imaging with the width that maximizes resolution.

    Authors do also a very good job describing, characterizing and designing experiments to deal with instabilities exhibited by the deformable mirror.

    One of the key aspects of the paper, that could be stressed more, is that including AO gives access to better-quality raw images that are then be used for standard reconstruction pipeline SIM processing. When aberrations are compensated, the illumination pattern closely matches what is expected by the SIM image formation model. Since these raw recorded images are sharper and closer to the actual assumption behind the SIM image reconstruction model, they will have a major positive impact in reducing artifacts that the inversion algorithm is returning. This is particularly evident in Figure 4.

    Major comments:

    • There is an overall reference in the manuscript of the novelty possible range of applications of using an upright microscope configuration. Examples mentioned are tissue-based imaging, access to whole-mount specimens for manipulation and electrophysiology. However, authors fail to present any such applications. There is not a single example presented which could not have been obtained with an inverted microscope. Could the authors provide an example where a water-dipping is used. Expanded samples could be one case, since the thickness of the gel makes it difficult to image with an inverted microscope. Another possible example would be to label the extracellular space and do shadow imaging of the tissue (SUSHI PMID: 29474910). ExM might be simpler to do as part of revising the manuscript than SUSHI.
    • On the main text it is described a 5-fold volumetric resolution, which is confusing since authors only mention lateral and axial resolutions. Their measurements correspond to a ~1.6-fold lateral improvement and ~1.7-fold axial improvement. These are however not the 95% of the achievable resolution theoretical maximum, as stated in p7 SI (2 fold increase of 282nm), but only the 80-85%. This point should be rephrased in the manuscript.
    • [OPTIONAL] p4 and related to figure 2, it would be important to report also measurements of beads with SIM but without AO, just as done for WF. Is there an improvement of using AO on SIM? This is reported for the fixed cells but not for the beads.
    • Figure 2, it is odd the comparison between WF+/- AO and SIM +/- AO are done using different cellular structures. Since wavelengths used are not the same it is difficult to interpret if there is any improvement of using AO on SIM compared to SIM without AO. Same questions arise as above, Is there an improvement of using AO on SIM?
    • "A significant benefit and uniqueness of the Deep3DSIM design is its upright configuration, whereas commercial SIM systems are built around inverted microscopes and are usually restricted to imaging thin samples, such as cultured cells." (p5) is not correct. The commercial DeepSIM module from CREST Optics can be mounted on an inverted microscope as well as image deep into tissue (see https://crestoptics.com/deepsim/ and application notes therein) and be used with essentially any objective. This point should be rephrased in the text.
    • Fig 3 reports the improvements of AO on SIM for imaging over 10um in tissue. What are the Zernike modes measured? Or how does the pupil look like before and after correction? It would be also good to report the Fourier amplitudes as done in Fig 2C as a quantitative measure of improvement. It would be good to point out the artifacts observed on the BP SIM image reconstruction (labelled with 3x, fringes are noticeable).
    • Many key details relating to image acquisition and AO correction are missing for all figures. How is the AO optimization implemented? Is it implemented via a genetic algorithm (progressive optimization of parameters) or using more clever strategies? Not clear if the optimization is implemented using images obtained with flat illumination or after SIM imaging/processing of a given dataset. How long does the AO optimization take? How sensitive to noise is the process? What metric do they use to estimate the sensorless AO correction? On pag12, they say "Fourier domain image metric" for measurements with fine details; otherwise, ISOsense when not high frequencies are present. Could the authors report the formula used to calculate the first metric? What do they consider to be low and high frequencies in this case? Is there a reason why ISOsense is not always used, or is there an automatic way to choose between the two? How many images were acquired for AO correction? Which samples were corrected with ISOsense and which ones with Fourier domain image metric? (see for example the detailed experimental reporting in the Supp Mat from Lin et al Nat Commun 2021).
    • Fig 4. Data presented for larval brain tissue is a very clear example of adding AO to image deep into tissue as the effect at ~130 cannot be understated. Here too, it would be also good to report the Fourier amplitudes as done in Fig 2C as a quantitative measure of improvement and possibly the SNR of reconstructed images. Having a way to quantitatively describe how much better are those images would be great. Also, what are the aberrations corrected? Can the wavefront or Zernike amplitude of the modes be reported? Same as for Fig 3, details about AO correction are missing.
    • [OPTIONAL] "It is worth noting that aberrations can differ across larger fields, and therefore, after applying an average correction, residual aberrations can still be observed in some regions of the AO-corrected bead images. However, the overall PSF shapes were still dramatically improved with AO compared to the equivalent without AO." This point is very interesting although not result either in the main text or in the SI is presented.
    • "As we found that the aberrations change approximately linearly in depth, we could measure the aberration in two planes and calculate the corrections in intermediate planes by interpolation, an approach which we termed "multi-position AO"." This is, personally, one of the major contributions of this work to the community. Unfortunately, it is not reported in detail. Not only for SIM but for imaging with WF or confocal, such linear change for aberrations with depth is not well known. Again, here the details of AO correction and image metrics are missing. To establish that for most thick biological structures 'aberrations change approximately linearly in depth' would be foundational to the widespread use of AO within standard imaging. Would it be possible for the authors to elaborate on this point and present detailed results? What is the error from measuring and correcting more than 2 planes? What is the error from just measuring and AO correcting at the deeper plane, i.e. from a single measurement? Authors could also show a case in which a linear assumption works nicely (or how well it works). For example, comparing an intermediate plane (or a plane beyond) imaged after AO optimization or after coefficient interpolation of the Zernike modes and compare it against correcting directly that plane.
    • The image of the cos-7 cell in metaphase, for Fig 5 is, however, very disappointing. See Fig 1 of Novak et al Nat Commun 2018 for an example of a single z-plane of a cell in metaphase. Having the possibility to correct for the entire 3D volume, I would expect amazing 3D volumes (movies and/or projections) associated with this imaging which are not presented.
    • In Figure 6, they use AO in remote configuration mode to allow imaging of live specimens. It needs to be clarified if this is an a priori characterization that is then kept fixed while recording in time. The last acquired volume of fig 6A and B have a higher amount of artifacts with respect to time 00:00. Are those artifacts due to lower SNR (maybe due to sample bleaching) or due to some change in the aberrations of the specimen?
    • "These results demonstrate that the remote focusing functionality of the system can be successfully applied for live 3D-SIM experiments, allowing four-dimensional acquisition while keeping the specimen stationary, thus avoiding the usual agitation and perturbations associated with mechanical actuation." Generally, this statement is true, but for the specific example shown of drosophila embryogenesis is it relevant? If they use piezo-driven Z-stack imaging with AO, does that lead to incorrect reconstructions or motion-induced artifacts? Related to the results shown in Fig 6, the fair comparison would be AO SIM vs SIM (without AO), not AO SIM vs AO WF.
    • When performing remote focusing, is the effective NA of the imaged plane changing with respect to the NA of the objective used at its focal plane?
    • [OPTIONAL] Did the authors run calculations to explore whether a commercial upright microscope could be used instead of their design? Are there any fundamental flaws that would make impossible using a commercial base? If not, could an AO SIM module be designed such that it adds on a commercial base? It would be important to discuss this point.

    Minor comments

    • Fig 2 lacks a color bar for D panels, which is in log scale. Authors should also show the Fourier transform along the z direction.
    • p4, "Such minor aberrations tend to be insignificant in conventional microscopy modalities such as widefield and confocal (Wang and Zhang, 2021). Therefore..." If optical aberrations are insignificant for single cells in widefield and confocal why do experiments here? These sentences should be rephrased to motivate better the experiments performed.
    • Imaged microtubules look abnormal, 'dotty' (figure 2) in both WF and SIM. See https://sim.hms.harvard.edu/portfolio/microtubules/ or Fig 1 of Wegel, et al Dobbie Sci Rep 2016, for better examples of continuous microtubule structures as imaged with SIM.
    • Is also the remote focusing performed via optimization of metrics similar to the one used for compensating aberrations?
    • Figure 2, the order of names on the top right of the panel should match the order of curves presented.
    • I value the efforts to improve open-source tools for system and AO control and GUI. And those tools seemed to have been modified for this work, although those modifications are not described. Would it be possible for the authors to describe those modifications?
    • Reported average values of the FWHM of imaged beads in 3D (p4) require also to report errors associated with those measurements.
    • Page 13, second paragraph states that "The results from chapter 3.2..." I believe that was a copy/paste from a thesis but should be corrected for a peer-reviewed publication, as there is no chapter 3.2.

    Referee Cross-Commenting

    The other two reviewers raise relevant and important points that would contribute to the overall improvement of the work. I think that authors should try to address most, if not all, of the comments as long as they don't require more than 3-6 months to get done.

    Significance

    General assessment:

    Although a very good and timely idea is presented the overall the manuscript still needs a lot of work. There is a lack of many key details of AO correction, all applications chosen could have been done in an inverted scope and some of the example images reported are suboptimal (Fig 2 and 5) that need further experimental work. Details and metrics, one example of the advantage of using an upright microscope and overall better examples of imaged cells could be provided.

    This work builds upon recent work of implementing AO for 3D SIM (Lin et al Nat Commun 2021) to propose to use a deformable mirror to perfrom AO as well as remote focusing in an upright microscope configuration.

    Audience: this work will be of interest for a specialized group of researchers, but it will contribute to the goal of adding AO tools to every microscope that will greatly impact the whole imaging community.

  3. Review Commons

    Note: This preprint has been reviewed by subject experts for Review Commons. Content has not been altered except for formatting.

    Learn more at Review Commons

    Referee #2

    Evidence, reproducibility and clarity

    The authors want to develop a structured illumination microscopy (SIM) system for deep tissue superresolution imaging. Here they have developed a SIM system based on the upright configration, and use deformable mirror to compensate for the relection index (RI) mismatch and improve the resolution and contrast in deep tissues. They also showed examples of SR imaging of COS-7 cells and Drosophila larval brains and embryos.

    However, I do have some concerns regarding the paper.

    1. The authors have provided an incomplete description of the structured illumination microscopy (SIM) reconstruction process. It is unclear whether the approach is based on 2D interference SIM configurations or 3D interference patterns. Furthermore, the specific algorithm utilized for image reconstruction has not been elucidated. Elaborating on these aspects is crucial as they significantly influence the interpretation of the resulting data.
    2. The authors have stated that sample-induced aberrations caused by RI inhomogeneities within the specimen is another major reason for causing artifacts generation. Literature has demonstrated that RI inhomogeneities can lead to non-local distortions in the grid pattern, which suggests that applying uniform reconstruction parameters across the entire image may not be viable. Traditional artifact remediation using the classical Wiener method is likely insufficient under these conditions (PMID: 33896197). The existing adaptive optics (AO) approach, which employs a deformable mirror (DM) alongside an sCMOS camera, is inadequate for tackling the issue at hand. Actually the assertion made in the paper that "aberrations change approximately linearly with depth" is seemingly contradicted by simulations referenced in the cited literature (PMID: 33896197). Consequently, it appears that the current methodology might only achieve a partial mitigation of the problems associated with spherical aberration resulting from RI mismatches. It is advisable, therefore, that the authors explicitly acknowledge this limitation in their manuscript to prevent any potential misinterpretation by readers.
    3. In Figure 2, the use of COS-7 cells, which are known for their relatively thin axial dimension, for the experiments raises an eyebrow. Notably, there are ample instances in existing research where both 2D-SIM and 3D-SIM, without the integration of adaptive optics, have yielded high-quality super-resolution images of structures such as tubulin and the endoplasmic reticulum. In addition, the authors did not present a direct comparison between BP-SIM and AO-SIM here. Without this comparative analysis, it remains ambiguous whether the enhancements in resolution and contrast and the reduction in artifacts can genuinely be attributed to the mitigation of spherical aberration. To clarify this, it would be beneficial for the authors to include side-by-side comparisons of these modalities to demonstrate the specific improvements attributed to AO-SIM.
    4. In Figures 3 and 4, the authors have illustrated the enhancements achieved through the application of AO. However, there is a discernible presence of hammer-stroke and honeycomb artifacts in pre-AO imaged data, which seem to originate from the amplification of the incorrectly moved out-of-focal background in the frequency domain. Various strategies have been previously suggested to address these specific artifacts, encompassing methods like subtracting background noise in the raw images or employing selective frequency spectrum attenuation techniques, such as Notch filtering and High-Fidelity SIM. To facilitate a more comprehensive understanding, I would recommend that the authors incorporate into their study a comparison that includes BP-SIM data that has undergone either background subtraction or frequency spectrum attenuation. This added data would enable a more complete evaluation and comparison regarding the merits and impact of their AO approach.

    Significance

    The authors want to develop a structured illumination microscopy (SIM) system for deep tissue superresolution imaging. Here they have developed a SIM system based on the upright configration, and use deformable mirror to compensate for the relection index (RI) mismatch and improve the resolution and contrast in deep tissues.

  4. Review Commons

    Note: This preprint has been reviewed by subject experts for Review Commons. Content has not been altered except for formatting.

    Learn more at Review Commons

    Referee #1

    Evidence, reproducibility and clarity

    Review, 3D SIM + AO, Wang and coworkers

    In this manuscript, Wang and coworkers report an upright 3D SIM system with adaptive optics (AO) correction. They demonstrate that AO improves imaging into thick 3D samples, including Drosophila larval brain. They also explore the use of remote focusing with their setup. The authors clearly demonstrate a gain with AO, and we are convinced that the microscope they build offers some utility over existing state of the art, particularly in samples thicker than a single cell. That said, we have concerns with the manuscript that we would like to see addressed before recommending publication:

    • Given the emphasis on super-resolution imaging deep inside a sample, we were surprised to see no mention of other forms of structured illumination that allow super-resolution imaging in samples thicker than a single cell. These include the 'spot-scanning' implementations of SIM that offer better imaging at depth by virtue of pinholes, and include MSIM, iSIM, and rescan confocal technologies. The two-photon / AO implementation of iSIM seems particularly germane, e.g. https://pubmed.ncbi.nlm.nih.gov/28628128/ Please consider citing these works, as they help place the existing work into context.
    • As we're sure the authors appreciate, besides aberrations, a major additional obstacle to 3D SIM in thick tissues is the presence of out-of-focus background. Indeed, this point was mentioned by Gustafsson in his classic 2008 paper on 3D SIM (https://pubmed.ncbi.nlm.nih.gov/18326650/): 'The application area of three-dimensional structured illumination microscopy overlaps with that of confocal microscopy, but the two techniques have different and complementary strengths. Structured illumination microscopy offers higher effective lateral resolution, because it concentrates much of the excitation light at the very highest illumination angles, which are most effective for encoding high-resolution information into the observed data, whereas confocal microscopy spreads out its illumination light more or-less uniformly over all available angles to form a focused beam. For very thick and compactly fluorescent samples, however, confocal microscopy has an advantage in that its pinhole removes out-of focus light physically. Structured illumination microscopy is quite effective at removing out-of-focus light computationally, because it is not subject to the missing-cone problem, but computational removal leaves behind the associated shot noise. Therefore confocal microscopy may be preferable on very thick and dense samples, for which the in-focus information in a conventional microscope image would be overwhelmed by out-of-focus light, whereas structured illumination microscopy may be superior in a regime of thinner or sparser samples.' This point is not mentioned at all in the manuscript, yet we are certain it is at least partially responsible for the residual image artifacts the authors mention. Please discuss the problem of out of focus light on 3D samples, particularly with an eye to the 'spot-scanning' papers mentioned above.
    • The authors use a water dipping lens, yet they image into samples that are mounted on coverslips, i.e. they use a dipping lens to image through a coverslip: see attached pdf for reference

    This almost certainly introduces spherical aberration, which the authors seem to observe: see attached pdf for reference

    We find this troubling, as it seems that in the process of building their setup, the authors have made a choice of objective lens that introduces aberrations - that they later correct. At the very least, this point needs to be acknowledged in the manuscript (or please correct us if we're wrong) - as it renders the data in Figs. 3-4 somewhat less compelling than if the authors used an objective lens that allowed correction through a coverglass, e.g. a water dipping lens with a correction collar. In other words, in the process of building their AO setup, the authors have introduced system aberrations that render the comparison with 3D SIM somewhat unfair. Ideally the authors would show a comparison with an objective lens that can image through a glass coverslip.

    • The authors tend to include numbers for resolution without statistics. This renders the comparisons meaningless in my opinion; ideally every number would have a mean and error bar associated with it. We have included specific examples in the minor comments below.
    • In Fig. 5, after the 'multipoint AO SIM', the SNR in some regions seems to decrease after AO: see attached pdf for reference

    Please comment on this issue.

    • Please provide timing costs for the indirect AO methods used in the paper, so the reader understands how this time compares to the time required for taking a 3D SIM stack. In a similar vein, the authors in Lines 213-215, mention a 'disproportionate measurement time' when referring to the time required for AO correction at each plane - providing numbers here would be very useful to a reader, so they can judge for themselves what this means. What is the measurement time, why is it so long, and how does it compare to the time for 3D SIM? It would also be useful to provide a comparison between the time needed for AO correction at each (or two) planes without remote focusing (RF) vs. with RF, so the reader understands the relative temporal contributions of each part of the method. We would suggest, for the data shown in Fig. 5, to report a) the time to acquire the whole stack without AO (3D SIM only); b) the time to acquire the data as shown; c) the time to acquire the AO stack without RF. This would help bolster the case for remote focusing in general; as is we are not sure we buy that this is a capability worth having, at least for the data shown in this paper.
    • Some further discussion on possibly extending the remote focusing range would be helpful. We gather that limitations arose from an older model of the DM being used, due to creep effects. We also gather from the SI that edge effects at the periphery of the DM was also problematic. Are these limitations likely non-issues with modern DMs, and how much range could one reasonably expect to achieve as a result? We are wondering if the 10 um range is a fundamental practical limitation or if in principle it could be extended with commercial DMs.

    Minor comments

    • The paper mentions Ephys multiple times, even putting micromanipulators into Fig. 1 - although it is not actually used in this paper. If including in Figure 1, please make it clear that these additional components are aspirational and not actually used in the paper.
    • The abstract mentions '3D SIM microscopes', 'microscopes' redundant as the 'm' in 'SIM' stands for 'microscope'.
    • 'fast optical sectioning', line 42, how can optical sectioning be 'fast'? Do they mean rapid imaging with optical sectinong?
    • line 59, 'effective imaging depth may be increased to some extent using silicone immersion objectives', what about water immersion objectives? We would guess these could also be used.
    • line 65 - evidence for 'water-dipping objectives are more sensitive to aberrations' ? Please provide citation or remove. They are certainly more prone to aberrations if used with a coverslip as done here.
    • 'fast z stacks' is mentioned in line 103. How fast is fast?
    • line 116 'we imaged 100 nm diameter green fluorescent beads'. Deposited on glass? Given that this paper is about imaging deep this detail seems worth specifying in the main text.
    • lines 127-130, when describing changes in the bead shape with numbers for the FWHM, please provide statistics - quoting single numbers for comparison is almost useless and we cannot conclude that there is a meaningful improvement without statistics.
    • In the same vein, how can we understand that remote focus actually improves the axial FWHM of the widefield bead? Is this result repeatable, or it just noise?
    • line 155, 'Because of the high spatial information...' -> 'Because of the high resolution spatial information...'
    • When quoting estimated resolution #s from microtubules (lines 158-163) similarly please provide statistics as for beads.
    • It seems worth mentioning the mechanism of AO correction (i.e. indirect sensing) in the main body of the text, not just the methods.
    • How long do the AO corrections take for the datasets in the paper?
    • Were the datasets in Fig. 2-4 acquired with remote focusing, or in conventional z stack mode? Please clarify this point in the main text and the figure captions.
    • It would be helpful when showing z projections in Figs. 3-5 to indicate the direction of increasing depth (we assume this is 'down' due to the upright setup, but this would be good to clarify)
    • line 174, 'showed significant improvements in both intensity and contrast after reconstruction' - we see the improvements in contrast and resolution, it is harder to appreciate improvements in intensity. Perhaps if the authors showed some line profiles or otherwise quantified intensity this would be easier to appreciate.
    • line 195 'reduced artefacts' due to AO. We would agree with this statement - the benefit from AO is obvious, and yet there are still artefacts. If the authors could clarify what these (residual) artefacts are, and their cause (out of focus light, uncorrected residual aberrations, etc) this would be helpful for a reader that is not used to looking at 3D SIM images.
    • Line 197, 'expected overall structure', please clarify what is expected about the structure and why.
    • Line 199, what is a 'pseudo structure'?
    • Fig. 4B, 'a resolution of ~200 nm is retained at depth', please clarify how this estimate was obtained, ideally with statistics.
    • Fig. 4D, please comment on the unphysical negative valued intensities in Fig. 4D, ideally explaining their presence in the caption. It would also be helpful to highlight where in the figure these plots arise, so the reader can visually follow along.
    • Line 245, 'rapid mitosis'. What does rapid mean, i.e. please provide the expected timescale for mitosis.
    • For the data in Fig. 6, was remote refocusing necessary?
    • What is the evidence for 'reduced residual aberrations', was a comparative stack taken without AO? In general we feel that the results shown in Fig. 6 would be stronger if there were comparative results shown without AO (or remote focusing).
    • Line 350, 'incorporation of denoising algorithms' - citations would be helpful here.
    • Line 411, 'All three were further developed and improved' - vague, how so?
    • Sensorless AO description; how many Zernike modes were corrected?
    • Multi-position aberration correction. Was the assumption of linearity in the Zernike correction verified or met? Why is this a reasonable assumption?
    • Fig. S1B is not useful; if the idea is to give a visual impression of the setup, we would recommend providing more photos with approximate distances indicated so that the reader has a sense of the scale of the setup. As is - it looks like a photograph of some generic optical setup.
    • SI pattern generation - 'the maximum achievable reconstruction resolution was only slightly reduced to about 95% of the theoretical maximum'. We don't understand this sentence, as the resolution obtained on the 100 nm beads is considerably worse than 95% of the theoretical maximum. Or do the authors mean 95% of the theoretical maximum given their pitch size of 317 nm for green and 367 nm for red? SI Deformable mirror calibration

    'spanning the range [0.1, 0.9]' - what are the units here?

    What are the units in Fig. S5C, S5D?

    It would be useful to define 'warmup' also in the caption of SI Fig. S6A. SI Remote Focusing, 'four offsets, {-5 mm, -2.5 mm, 2.5 mm, 5 mm}...' are the units mm or um? '...whereas that of the 10 beads was...' here, do the authors mean the position of the beads derived from the movement of the piezo stage, as opposed to the remote focusing? The authors refer to the 'results from Chapter 3.2'. What are they talking about? Do they mean a supplementary figure, or earlier supplementary results? In general, we found the discussion in this paragraph difficult to follow. Supplementary Fig. 9 seems to be not referred to anywhere in the text.

    • Since the paper emphasizes 3D SIM, OTFs along the axial direction would also be useful to show, in addition to the lateral OTFs shown in Fig. 2D.
    • When the sample is moved by the piezo, the axial phase of the 3D-SIM illumination pattern is stable as the sample is scanned through the illumination pattern. When remote focusing is performed, the sample is always stable so the axial phase of the 3D-SIM illumination pattern is presumably changing with remote focusing. Can the authors clarify if the 3D SIM illumination pattern is scanned when remote focusing is applied, or is the intensity pattern stable in z?
    • In Supplementary Fig. 9, primary spherical is referred to twice, both at index 11 and 22. The latter is presumably secondary spherical?
    • we do not understand the x axis label, in Fig. S4D, is it really [0, 50, 50, 50] as written? see attached pdf for reference

    Referee Cross-Commenting

    I don't have much to add; the other reviewers raise good points and I think it would be good if the authors could respond to their feedback in a revised manuscript.

    Significance

    Nearly all fluorescence images deteriorate as a function of depth. Methods to ameliorate this depth-dependent degradation are thus of great practical value, as they improve the information content of images and thus (hopefully) biological insight. In this work, the authors develop a method to improve super-resolution imaging (3D SIM) at depth, by combining it with adaptive optics.

  5. Version published to 10.1101/2023.12.21.572818v2 on bioRxiv
  6. Version published to 10.1101/2023.12.21.572818v1 on bioRxiv