These authors contributed equally to this work
Imaging neuronal activity with high and homogeneous spatial resolution across the field-of-view (FOV) and limited invasiveness in deep brain regions is fundamental for the progress of neuroscience, yet is a major technical challenge. We achieved this goal by correcting optical aberrations in gradient index lens-based ultrathin (≤500 µm) microendoscopes using aspheric microlenses generated through 3D-microprinting. Corrected microendoscopes had extended FOV (
The amount of information carried by neural ensembles and the impact that ensemble activity has on signal propagation across the nervous system and on behavior critically depend on both the information and tuning properties of each individual neuron and on the structure of correlated activity, either at the level of correlations between each pair of neurons or at the whole network level (
A critical barrier to progress is the lack of availability of microendoscopic devices with small cross-sections that maintain cellular resolution across a large FOV, to allow high-resolution and high SNR two-photon population imaging on a large number of neurons while minimizing tissue damage. Preserving short- and long-range connectivity is fundamental to study physiological network dynamics in brain circuits. For example, severing thalamocortical fibers alters low-frequency spontaneous oscillations in the thalamocortical loop (
Here, we report the design, development, and characterization of a new approach to correct aberrations and extend the FOV in ultrathin GRIN-based endoscopes using aspheric lenses microfabricated with 3D micro-printing based on two-photon lithography (TPL) (
Four types (type I-IV) of
(
For each type of GRIN rod used in the
(
(
(
(
(
The simulated focal length in the absence and presence of the corrective microlens for the four different types of microendoscopes is reported in
Corrective lenses were experimentally fabricated using TPL (
(
(
(
(
(
(
(
(
To visualize the profile of fluorescence intensity across the whole diameter of the FOV for both uncorrected and corrected probes, we used a subresolved thin fluorescent layer (thickness: 300 nm) as detailed in
To validate
To establish a quantitative relationship between the improved optical properties of
(
Comparison between data obtained with uncorrected and corrected microendoscopes in silico and in vivo.
(
(
Pairwise correlation between nearby neurons (distance between the center of neurons <20 µm) should not vary with the radial distance because in our simulations this value was constant across neurons. However, we found an artefactual increase of correlation strength with the radial distance of neuronal pairs in uncorrected endoscopes due to the cross-contamination of activity at different points generated by the larger and aberrated PSF without the corrective lens. In contrast, correlation strength remained constant in
We experimentally validated the results of the simulations performing functional imaging in the VPM using type II
To evaluate if the segmentation method could affect these results, we compared the quality of the manual segmentation method used in previous experiments with that of a standard automated algorithm (e.g. CaImAn, [
We evaluated the effect of aberration correction on the output of the analysis in the simulated and experimental data shown in
We then focused our attention on the VPM, a key region which relays somatosensory (whisker) information to the barrel field of the primary somatosensory cortex (S1bf) through excitatory thalamocortical fibers (
As an important control experiment, we first confirmed that the ultrathin GRIN lenses that we used in our study (diameter ≤500 µm) did not significantly damage anatomical thalamocortical and corticothalamic connectivity, a difficult task to achieve with larger cross-section GRIN lenses or with chronic optical windows (
(
We then used
(
We investigated how neuronal activity was modulated by an important behavioral variable: whether the mouse was whisking or not. We considered neuronal activity both at the single-cell and population level. We quantified the content of mutual information about whisking state (whether or not the mouse was whisking; shortened to whisking information hereafter) based on the fluorescence signals extracted from individual neurons (
(
Values of cell-specific and population information analysis.
Spatial map of neurons encoding whisking information in 12 out of the total 24 analyzed time series. The pseudocolor scale shows significantly informative neurons (see Materials and methods).
We also considered the redundancy and synergy of whisking information carried by pairs of simultaneously recorded nearby (distance between neurons < 20 µm) neurons. This analysis is important because how pairwise correlations shape the redundancy and synergy of information representation is fundamental to the understanding of population codes (
We finally turned to analyzing the properties of firing at the level of the whole population recorded in the FOV. We applied non-negative matrix factorization (NMF) (
Improving optical performances in ultrathin (diameter ≤0.5 mm) microendoscopes with built-in optical elements is a major technical challenge. Since the insertion of the probe irreversibly damages the tissue above the target area, reducing the size of the probe and consequently its invasiveness is of utmost importance when imaging deep brain regions. In this study, we designed, developed, characterized, and successfully validated a new approach to correct aberrations in ultrathin GRIN-based endoscopes using aspheric lenses microfabricated with 3D micro-printing based on TPL (
Corrective lenses were fabricated on glass coverslips, which were aligned and assembled with the GRIN rod to form an aberration-corrected microendoscope. This optical design resulted in improved axial resolution and extended effective FOV (
Aberration correction in GRIN microendoscopes can be achieved using adaptive optics (AO) (
Using synthetic calcium data, we demonstrated that the improved optical properties of
Studying neuronal population codes requires the measurement of neuronal population activity with high-precision, large SNR, and without introducing artificial bias on the activity of individual neurons and the measures of relationship between them, such as pairwise correlations. In particular, pairwise correlations are thought to be fundamental for information coding, signal propagation, and behavior (
Corrected endoscopes were characterized by a curved FOV. In the case of type II corrected endoscopes, the change in the z coordinate in the focal plane was up to ~75 µm (
We used the unique features of the
One important area of future development for
In summary, we developed a new methodology to correct for aberrations in ultrathin microendoscopes using miniaturized aspheric lenses fabricated with 3D printing based on TPL. This method is flexible and can be applied to the GRIN rods of different diameters and lengths that are required to access the numerous deep regions of the mammalian brain. Corrected endoscopes showed improved axial resolution and up to nine folds extended effective FOV, allowing high- resolution population imaging with minimal invasiveness. Importantly, we demonstrated that
Reagent type | Designation | Source or reference | Identifiers | Additional information |
---|---|---|---|---|
Strain, strain | C57BL/6J | Charles River | RRID: | |
Genetic reagent | B6;C3-Tg(Scnn1a- | The Jackson Laboratory | RRID: | |
Recombinant | pAAV.Syn.Flex. | Penn Vector Core | RRID: | |
Recombinant | pGP-AAV-syn- | Addgene | RRID: | |
Recombinant | AAV pCAG- | Penn Vector Core | RRID: | |
Recombinant | AAV.CaMKII0.4. | Penn Vector Core | RRID: | |
Commercial | Kwik-Cast | World Precision Instruments | Cat# KWIK-CAST | |
Commercial | Sylgard Silicone | Dow Inc | Cat# Sylgard 164 | |
Commercial | Norland Optical | Norland | Cat# NOA 63 | |
Commercial | GRIN lens | Grintech | Cat# NEM-050- | |
Commercial | GRIN lens | Grintech | Cat# NEM-050- | |
Commercial | GRIN lens | Grintech | Cat# GT-IFRL-035- | |
Commercial | GRIN lens | Grintech | Cat# NEM-035- | |
Chemical compound, | bisBenzimide H 33342 | Sigma-Aldrich | Cat# B2261; CAS: 23491-52-3 | |
Chemical compound, | Red Retrobeads | LumaFluor Inc | Red Retrobeads | |
Software, algorithm | Zemax OpticStudio 15 | Zemax | ||
Software, algorithm | MATLAB R2017a | Mathworks | RRID: | |
Software, algorithm | GraphPad PRISM | GraphPad PRISM | RRID: | |
Software, algorithm | ImageJ/Fiji | Fiji | RRID: | |
Software, algorithm | NoRMCorre | |||
Software, algorithm | CaImAn | |||
Software, algorithm | ||||
Software, algorithm | LIBSVM | |||
Software, algorithm | Information | N/A | ||
Software, algorithm | Software used in this | |||
Software, algorithm | Software to compute | |||
Other | Basler ace camera | Basler AG | Cat# acA800-510um | |
Other | Optical encoder | Broadcom | AEDB-9140-A13 | |
Other | Zortrax M200 3D printer | Zortrax | M200 | |
Other | Z-ULTRAT 3D | Zortrax | Z-ULTRAT | |
Other | Arduino Uno | Arduino | Arduino Uno |
Experimental procedures involving animals have been approved by the Istituto Italiano di Tecnologia Animal Health Regulatory Committee, by the National Council on Animal Care of the Italian Ministry of Health (authorization # 1134/2015-PR, # 689/2018-PR) and carried out according to the National legislation (D.Lgs. 26/2014) and to the legislation of the European Communities Council Directive (European Directive 2010/63/EU). Experiments were performed on adult (8–14 week old) mice. C57BL/6J mice (otherwise called C57, Charles River #000664, Calco, IT) were used in
Simulations were run with OpticStudio15 (Zemax, Kirkland, WA) to define the profile of the aspheric corrective lens to be integrated in the aberration-corrected microendoscopes, with the aim to achieve: (i) a full-width half maximum (FWHM) lateral resolution <1 µm at the center of the FOV; (ii) a FWHM axial resolution below <10 µm; (iii) a working distance between 150 µm and 220 µm into living brain tissue. The wavelength used for simulations was λ = 920 nm. The surface profile of corrective aspheric lenses was described in
Since GRIN lenses have intrinsic spherical aberration, the optimization for the shape of the corrective lenses started with the profile of a Schmidt corrector plate (
To evaluate the effect of corrective lenses on the 3D image space (
The optimized aspheric lens structure obtained with simulations was exported into a 3D mesh processing software (MeshLab, ISTI-CNR, Pisa, IT) and converted into a point cloud dataset fitting the lens surface (with ~300 nm distance among first neighborhood points). Two-photon polymerization with a custom set-up (
For fast generation of multiple lens replicas, a molding (
Optical characterization of
Adeno-associated viruses (AAVs) AAV1.Syn.flex.GCaMP6s.WPRE.SV40, AAV1.CAG.Flex.eGFP.WPRE.bGH, AAV1.CaMKII0.4.Cre.SV40 were purchased from the University of Pennsylvania Viral Vector Core. AAV1.Syn.flex.GCaMP7f.WPRE.SV40 was purchased from Addgene (Teddington, UK) Animals were anesthetized with isoflurane (2% in 1 L/min O2), placed into a stereotaxic apparatus (Stoelting Co, Wood Dale, IL) and maintained on a warm platform at 37°C. The depth of anesthesia was assessed by monitoring respiration rate, heartbeat, eyelid reflex, vibrissae movements, reactions to tail and toe pinching. 2% lidocaine solution was injected under the skin before surgical incision. A small hole was drilled through the skull and 0.5–1 µl (30–50 nl/min, UltraMicroPump UMP3, WPI, Sarasota, FL) of AAVs containing solution was injected at stereotaxic coordinates: 1.4 mm posterior to bregma (P), 1 mm lateral to the sagittal sinus (L), and 1 mm deep (D) to target the hippocampal CA1 region; 1.7 mm P, 1.6 mm L, and 3 mm D to target the VPM. Co-injection of AAV1.Syn.flex.GCaMP6s.WPRE.SV40 and AAV1.CaMKII0.4.Cre.SV40 (1:1) was performed to express GCaMP6s in hippocampus CA1 pyramidal cells of C57 mice (
To evaluate thalamocortical (TC) and corticothalamic (CT) anatomical connectivity (
For experiments in anesthetized conditions (
For experiments in behaving mice (
In
Deeply anesthetized animals were transcardially perfused with 0.01 M PBS (pH 7.4) followed by 4% paraformaldehyde. Brains were post-fixed for 6 hr, cryoprotected with 30% sucrose solution in 0.1 M PBS, and serially cut in coronal sections (thickness: 40–50 µm) using a HM 450 Sliding Microtome (Thermo Fisher). Sections were counterstained with Hoechst (1:300, Sigma–Aldrich, Milan, IT), mounted, and coverslipped with a DABCO [1,4-diazobicyclo-(2,2,2)octane]-based antifade mounting medium. Fluorescence images were acquired with a Leica SP5 inverted confocal microscope (Leica Microsystems, Milan, IT).
For the evaluation of the anatomical connections between VPM and S1bf, mice were perfused after 10 days from the injection and GRIN lens implantation. 50 µm thick coronal brain slices were cut, counterstained with Hoechst (1:300, Sigma–Aldrich, Milan, IT), and mounted with an Antifade Mounting Medium (Vectashield, Burlingame, CA). Confocal images were acquired with a Nikon Eclipse scope (Nikon, Milan, IT).
In
Neural spiking activity was simulated as the sum of Poisson processes. Each neuron was assigned with a mean spiking rate (
The size and the resolution of the simulated FOV were set to 500 × 500 µm2 and 2.5 µm/pixel, respectively. The resolution was adjusted according to the changes in the magnification factor (estimated from experimental data
To generate the synthetic
Excitation volumes were scanned along the imaging focal surface (or surfaces for uncorrected microendoscopes), such that their axial direction was always orthogonal to the imaging focal surface(s). All the voxels falling within the excitation volumes contributed to the signal of the corresponding pixel in the FOV, resulting in one of the following possible three conditions:
If the pixel was in the edge of the FOV (radial distance >250 µm), its signal was randomly sampled from a normal distribution, with mean and standard deviation estimated from experimental data. For this and the following conditions, we selected the best fitting distribution for the signal mean through log-likelihood maximization across four alternative models: a normal, a gamma, a log-normal distribution, and a Gaussian mixture model. Dark noise mean was best fitted by a Gaussian mixture model (component 1: proportion = 0.37; mean = 137.48; sd = 48.96; component 2: proportion = 0.63; mean = 126.83; sd = 5.02, Log-likelihoods: Normal = −3.3E5, Gamma = −3.0E5, Log-normal = −2.9E5, Gaussian mixture model = −2.6E5), while the standard deviation of the dark noise depended on the dark noise mean in a linear way (p0 = −175.39, p1 = 1.57). The simulated dark noise was generated with the mean randomly sampled from the Gaussian Mixture Modeling (GMM) distribution and the standard deviation linearly dependent from the mean.
If the pixel was in the central part of the FOV (radial distance ≤250 µm) but no neurons were within the excitation volume, the pixel signal was randomly sampled from a normal distribution with mean and standard deviation estimated from experimental data. The mean intensity of pixels that were neither in the edges nor belonging to ROIs were fitted using a lognormal distribution (mean = 5.43, sd = 0.36. Log-likelihoods: Normal = −1.46E6, Gamma = −1.39E6, Log-normal = −1.37E6, Gaussian mixture model = −1.39E6) and the best linear fit between the squared root of the mean intensity and the intensity sd was computed (p0 = −162.55, p1 = 18.28). Simulated noise in the FOV was generated as Gaussian noise with mean randomly sampled from the lognormal distribution and sd linearly dependent from the squared root of the mean.
If the pixel was in the central part of the FOV (radial distance ≤250 µm) and at least one neuron was in the excitation volume(s), each voxel in the excitation volume(s) was assigned either Gaussian noise (estimated as in the previous condition. Log-likelihoods: Normal = −2.70E5, Gamma = −2.60E5, Log-normal = −2.58E5, Gaussian mixture model = −2.59E5) in case no neurons were in that voxel, or the fluorescence intensity of the neuron sampled by that voxel. In case a neuron was contained in a voxel, Gaussian noise was also added to the neuron signal. The mean of the added Gaussian noise was zero, while the sd was proportional to the square root of the mean intensity of the voxel, with the coefficients estimated from a linear fit between the square root of the mean intensity and the intensity standard deviation of pixels assigned to ROIs in experimental data (p0 = −132.44, p1 = 16.94). The activity of all the voxels falling within the excitation volume(s) was then averaged to obtain the pixel’s fluorescence intensity. The intensity of each pixel signal was finally modulated as a function of the radial position within the FOV, accordingly to the optical characterization of corrected and uncorrected microendoscopes using the radial intensity obtained imaging the subresolved fluorescent layer (
In simulations, the imaging rate of the
We segmented simulated time series using two approaches: an automated procedure that we developed to resemble manual segmentation, and a standard automated approach (CaImAn [
For the CaImAn segmentation, we tested different values of the SNR parameter (0.25, 0.5, 1, 1.5, 2), that represented a lower SNR threshold for a ROI to be kept in the final segmentation.
Values are expressed as mean ±sem, unless otherwise stated; the number of samples (n) and p values are reported in the Figure legends or in the text. No statistical methods were used to pre-determine sample size. All recordings with no technical issues were included in the analysis and blinding was not used in this study. Statistical analysis was performed with MATLAB software (Mathworks, Natick, MA) and GraphPad Prism software (GraphPad Software, San Diego, CA). A Kolmogorov-Smirnov test was run on each experimental sample to test for normality and to test the equality of the distributions in
Three mice were unilaterally injected with AAV-eGFP and red retrobeads after tissue aspiration and then implanted with the endoscope. Three mice were injected with AAV-GFP and red retrobeads without tissue aspiration and they were not implanted (controls). Three confocal images were acquired from fixed slices for each hemisphere at different focal planes (minimal distance between planes: 20 µm). Images in the red and green acquisition channels were blurred with a Gaussian filter (sigma = 2 µm) and binarized with a triangle thresholding method. S1bf was manually identified using anatomical cues from Hoechst labeling in each sample. To quantify the amount of preserved TC and CT connections within a given area of S1bf, we computed the fraction of pixels showing suprathreshold pixel intensity out of the total of pixel of the chosen area. A single value for each sample, obtained by averaging between confocal images of different FOVs, was used to run the one-tailed Mann-Whitney test for the different acquisition channels (
A regular fluorescent grid spanning the FOV was imaged in order to evaluate the distortion in the FOV. The number of pixels necessary to span 10 µm in the x and y direction was measured as a function of the distance from the FOV center. A magnification factor which varied along the radial directions was evaluated by computing the ratio between the measured number of pixels in the distorted (microscope objective coupled with GRIN-lens-based microendoscope) and undistorted (microscope objective alone) conditions. The estimated magnification factor (from x and y directions) was fitted using a quadratic curve (corrected: p0 = 0.76, p1 = −6.24E-04, p2 = 1.95E-05, norm of residuals = 1.78; Uncorrected: p0 = 0.73, p1 = −2.91E-04, p2 = 8.11e-06, norm of residuals = 0.24). The magnification factor was used to correctly calibrate experimental measurements in
To measure the PSF as a function of the radial position within the FOV, z-stacks of subresolved fluorescent beads (diameter: 100 nm) were taken at different distances from the optical axis. Intensity profiles obtained from sections in the x, y, z directions of the PSFs were fitted with Gaussian curves and their FWHM was defined as x, y, and z resolution, respectively. Lateral resolution was calculated as the average of x and y resolution. Axial resolution coincided with the z resolution. When, due to aberrations in the lateral portion of the FOV, the intensity profile in the z direction was better fitted with a sum of two Gaussian curves instead of a single one, the axial resolution was defined as the axial distance between the two peaks of the best fitting curves. For each group of measurements at a specific distance from the optical axis, outliers were identified using the Rout method (GraphPad Software, San Diego, CA) and excluded from the data. Mean and standard deviation of resolutions were plotted against radial distance (
Experiments in VPM were analyzed with a customized graphical user interface (GUI) in MATLAB (version R2017a; Mathworks, Natick, MA). The GUI enabled the loading of data saved from the microscope acquisition software, included the motion correction algorithm (NoRMCorre) described in
For synthetic data in
For experimental data in
For both synthetic and experimental data, the SNR was defined as:
For synthetic data, we computed the correlation between the calcium activity of each segmented ROI and the ground truth calcium activity of neurons contributing to that ROI. In case more neurons were merged during the automated segmentation that we developed, we sorted the merged neurons for decreasing correlation with the corresponding ROI. We defined ‘source neuron’ the neuron with highest correlation with the ROI (
For each pair of nearby ROIs (distance between ROIs centers < 20 μm), we computed pairwise correlation of the extracted calcium activity as a function of the radial distance of the ROIs pair. We defined the radial distance of each pair as the distance between the FOV’s center and the center of the segment connecting the two ROIs’ centers. To measure changes in pairwise correlations exclusively caused by changes in imaging resolution, we normalized the pairwise correlations of nearby ROIs by subtracting the value of pairwise correlations between distant ROIs (distance >60 μm) placed at the same radial distance. We then fitted the normalized pairwise correlation as a function of radial distance using linear regression.
Videos of whisker movements were binarized with the Ridge Detection plugin of ImageJ/Fiji in order to individuate pixels corresponding to whiskers. Videos were then processed in MATLAB (Mathworks, Natick, MA) to extract the whisker mean angle. To this aim, all whiskers were fitted with straight lines and for each frame the mean angle of all the lines was calculated with respect to the horizontal direction of the FOV. Once the mean angle of the imaged whiskers was calculated for each frame, this signal was processed with a moving standard deviation over a 400 ms window and a Gaussian filter over a 50 ms window. Whisking and no whisking periods were identified by binarizing the mean whisker angle with a temporal and amplitude threshold. While the temporal threshold was fixed at 200 ms, the amplitude threshold was extracted by manually identifying whisking periods in ~1/10th of the full-length video and using this manual classification to find the best amplitude threshold with a ROC analysis. Temporal gaps between whisking periods shorter than 0.5 s were considered whisking periods, and linear interpolation was used to obtain the whiskers mean angle in frames in which less than four whiskers were detected.
For the analysis of pupil diameter, movies were analyzed with MATLAB (Mathworks, Natick, MA). Each frame was thresholded with the Otsu’s method and the region corresponding to the pupil was approximated with an ellipse. The length of the major axis of the ellipse was considered the pupil diameter. Linear interpolation was used for frames in which the pupil was not properly detected.
Detection of locomotion periods was performed using a threshold criterion on the wheel speed (
Four behavioral states were defined in
To compare the amplitude of calcium activity across behavioral states, the deconvolved activity of each ROI was averaged in each of the three states (Q, W and WL).
To measure whether calcium activity was further modulated by arousal, we discretized the pupil size in ten bins and measured the distribution of the average calcium activity for each behavioral state, separately.
For information theoretical analyses, we used the MATLAB toolbox provided in
We sorted ROIs according to their information content and fitted the distribution of information content of individual ROIs using a double exponential function, where the information carried by the
We used the R (
To compute information carried by pairs of neurons, we considered only pairs of nearby neurons (distance between neurons < 20 μm). We computed the amount of synergistic information carried by the neuronal pair as a function of the pairwise correlation between neurons and as a function of the pair’s radial distance. Data in
To compute information about the whisking state from a large population of neurons, we first decoded the whisking state from the single-trial population activity, and then we expressed the decoding performance as mutual information contained in the confusion matrix as in Equation (11) of
To compute NMF (
Sparseness:
Whisking modulation index (WMI):
Spatial spread: we defined as spatial spread the shortest path that connected the ten ROIs with highest weights (or all the ROIs in case a module was composed by fewer ROIs).
To compare pairs of modules, we computed the following similarity measures (
The Jaccard index is the fraction between the number of ROIs belonging to both modules and the total number of ROIs composing the two modules (without repetitions). It ranges between 0 and 1 and assumes value 0 if two modules do not share common ROIs, value one if two modules are composed by exactly the same ROIs. The Jaccard index does not take into account the weights of ROIs in the modules.
Cosine similarity =
For both the Jaccard index and Cosine similarity, we computed null distributions. Specifically, for each NMF factorization we reassigned ROIs randomly within each module by shuffling their weights. We did not consider modules composed by single ROIs.
The datasets shown in
The software used in this paper to generate artificial
We thank M Dal Maschio for discussion at an initial stage of the project, F Nespoli for preliminary analysis, and B Sabatini and A Begue for critical reading an early version of this manuscript. We thank DS Kim and the GENIE project for the constructs Addgene viral prep # 100845-AAV1 and #104492-AAV1, H Zeng for the construct Addgene viral prep # 51502-AAV1, and JM Wilson for the construct Addgene viral prep # 105558-AAV1. This work was supported by an IIT interdisciplinary grant and in part by ERC (NEURO-PATTERNS), NIH Brain Initiative (U01 NS090576, U19 NS107464, R01NS109961), FP7 (DESIRE), MIUR FIRB (RBAP11 × 42L), and Flag-Era JTC Human Brain Project (SLOW-DYN). CL acknowledges support from KAUST under baseline funding BAS/1/1064-01-01.
No competing interests declared
Data curation, Formal analysis, Validation, Investigation, Visualization, Methodology, Writing - original draft
Data curation, Software, Formal analysis, Validation, Investigation, Visualization, Writing - original draft
Data curation, Software, Formal analysis, Validation, Investigation, Visualization, Writing - original draft
Data curation, Software, Formal analysis, Validation, Investigation, Visualization, Methodology, Writing - original draft
Data curation, Software, Formal analysis, Validation, Investigation, Visualization, Methodology
Validation, Investigation, Methodology
Formal analysis, Supervision, Validation
Formal analysis, Visualization, Writing - original draft
Validation, Methodology
Methodology
Formal analysis, Supervision, Methodology, Writing - original draft, Writing - review and editing
Conceptualization, Supervision, Funding acquisition, Methodology, Writing - original draft, Project administration, Writing - review and editing
Conceptualization, Resources, Supervision, Funding acquisition, Writing - original draft, Project administration, Writing - review and editing
Animal experimentation: Experimental procedures involving animals have been approved by the Istituto Italiano di Tecnologia Animal Health Regulatory Committee, by the National Council on Animal Care of the Italian Ministry of Health (authorization # 1134/2015-PR, # 689/2018-PR) and carried out according to the National legislation (D.Lgs. 26/2014) and to the legislation of the European Communities Council Directive (European Directive 2010/63/EU).
Supplementary Table 1: Parameters for the fabrication of corrective lenses. Coefficients used in
The datasets shown in Figures 1-7 and corresponding figure supplements are available at:
The following dataset was generated:
In the interests of transparency, eLife publishes the most substantive revision requests and the accompanying author responses.
In this revised manuscript, the authors demonstrated an extended field-of-view (eFOV), thin (<500μm) microendoscopes by adding aberration correcting microlenses fabricated by TPL. The performance of the microlens corrected eFOV endoscopes was evaluated with simulations on synthetic calcium data and in vivo activity imaging. In addition, the eFOV microendoscopes were used to investigate VPM activity correlated with locomotion, whisker movement, and pupil diameter. Cell-specific encoding in VPM during these behaviors is further scrutinized with statistical and machine learning methods. The presentation of the methods and the results are clear. The methods are useful and practical for increasing the FOV of ultrathin microendoscopes, removing one of the current limitations of small GRIN lens applications. The authors significantly improved the manuscript based on the reviewer's comments.
Thank you for submitting your article "Extended field-of-view ultrathin microendoscopes for high-resolution two-photon imaging with minimal invasiveness" for consideration by
The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.
We would like to draw your attention to changes in our revision policy that we have made in response to COVID-19 (https://elifesciences.org/articles/57162). Specifically, when editors judge that a submitted work as a whole belongs in
Our expectation is that the authors will eventually carry out the additional experiments and report on how they affect the relevant conclusions either in a preprint on bioRxiv or medRxiv, or if appropriate, as a Research Advance in
Your manuscript describes the development and use of microfabricated corrector lenses for improving GRIN lens imaging. These correction lenses increase the usable field-of-view (FOV) of a given GRIN lens, and improves the resolution. The approach was validated in a variety of samples. While the reviewers agree that the work will be potentially high impact, they feel that revisions with additional analyses would significantly improve the paper.
In this manuscript, the authors demonstrated an extended field-of-view (eFOV), thin (<500μm) microendoscopes by adding aberration correcting microlenses fabricated bt TPL. Performance of the microlens corrected eFOV endoscopes was demonstrated with simulations on synthetic calcium data and in vivo activity imaging. In addition, the eFOV microendoscopes were used to investigate VPM activity correlated with locomotion, whisker movement, and pupil diameter. Cell-specific encoding in VPM during these behaviors are further scrutinized with statistical and machine learning methods. The presentation of the methods and the results are clear. The methods are useful and reasonably practical for increasing the FOV of ultrathin microendoscopes, which is a major limitation for small GRIN lenses. Overall, I recommend publication of the paper in
1) For the design and fabrication of the eFOV-microendoscope:
a) The authors presented four types of GRIN lenses (Type I-IV). Four different models of the GRIN lens are shown in subsection “Corrective lens manufacturing and microendoscope assembly”. It will probably help if the authors can explicitly indicate the Types with the model number.
b) It is claimed that the experimental result is similar to ray-trace simulation as shown in Figure 2. However, there is no comparison to support the claim. The high-order coefficient in the gradient profile in GRIN lens is usually not included in the optical model used for ray-trace. High-order aberration will thus not be reflected in the simulation. It would be important to know the difference between the simulation result and the experimental result. For example, the authors can add experimental data in Figure 2 or add simulation data in Figure 3.
c) For all the experiments, is it performed with the 2-photon polymerized lens with resin or with lens replica using a molded UV-curable adhesive? Is the aberration correction performance different from these two fabrication methods? Which material is the ray-tracing simulation based on?
d) What is the yield of the lens fabricated? How is the cured UV-curable adhesive detached from the PDMS mold? During imaging sessions, was there any damage on the lens?
e) How does the correction microlens affect the focal length of the GRIN lens. In other words, what is the allowable change in working distance before and after the correction?
f) There appears to be a much larger field curvature with the corrected GRIN lens. It would be helpful to discuss the impact of this on functional imaging.
g) Instead of AO with an active element (e.g., SLM), one could also put a fixed lens with the desired curvature in the place of the SLM. Compared to the microlens approach, one needs to modify the excitation path somewhat, but the lens can be fabricated using standard techniques and may be more accessible to most research labs. It might be informative to the readers to comment on this alternative approach.
2) For the data taken with the eFOV-microendoscope, there is ~2x improvement in FOV for type III lens (Figure 3G). However, in Figure 3—figure supplement 5C, the images before and after correction for type III lens is almost the same.
3) For the data simulation and analysis:
a) Analysis was shown to correlate whisking, locomotion and cell-specific encoding. Figure 6C and 6D show that locomotion is usually correlated with a higher ΔF/F. Could this be a result of the motion artifact during locomotion from the head-constraint mouse?
b) To simulate synthetic calcium data, different models (Gaussian Mixture, lognormal, etc.) are used to obtain the pixel intensity value as described in subsection “Generation of fluorescence time series”. It is mentioned that the model is obtained by fitting the experimental data. What criteria (e.g. comparing likelihood) is used to choose those specific models? What other distribution is tested to choose the best-fit model?
c) As mentioned in subsection “Analysis of field distortion and calibration of the pixel size”, the author used a gaussian profile to define axial resolution. It is probably better to fit with a Lorentzian function I(z)=〖(1/(1+〖(z/zR)〗^2 ))〗^2 for 2-photon excitation, where zR is the Rayleigh range of the gaussian focus. Not sure how large the difference will be between these two fits.
d) Figure 4D. It is not clear to me why the ROI number is smaller for the corrected lens than the uncorrected when the peak SNR value is small (e.g., less than 10). Can the authors add some explanations?
e) The authors claim that "pairwise correlation increased as a function of the radial distance of the pair in uncorrected probes compared to corrected ones." Figure 4M does not seem to show noticeable increase as a function of the radial distance for the uncorrected lens (left panel). Rather, it seems to show a noticeable decrease as a function of the radial distance for the corrected lens (right panel). Therefore, the statement does not seem to be supported by Figure 4M. This is confusing, and some explanation is needed.
4) There is a typo in the first sentence of subsection “Fabrication of eFOV-microendoscopes”. I believe it is (Figure 3—figure supplement 1C) instead of (Figure 1C).
In this paper, the authors describe the creation and use of microfabricated corrector lenses for improving GRIN lens imaging. GRIN lenses are one of a few technologies (and the only one that is "mature") that allow for high resolution optical imaging in brains at depths beyond ~1.5 mm. These correction lenses a) increase the usable field-of-view (FOV) of a given GRIN lens, and b) improve the fidelity of signals extracted from regions-of-interest (ROIs), by decreasing the point spread function (PSF) volume. Both are important for high fidelity recording of neuronal populations. The paper is well written, and I find their demonstrations compelling. I think the work is of broad interest, and suitable for publication in
My comments below are mainly critical, but I note that I think the paper overall is very good, and thorough, and I have high enthusiasm.
Given that the surface profiles begin their optimization from an analytical expression, It would be nice if the authors could comment on the likely maximal possible correction possible – that is on a GRIN lens with a GRIN relay of significant pitch, do they think microlenses would be similarly manufacturable or effective?
In the simulated imaging data, especially the figure, and SNR conclusions drawn from that, I think it is important to more clearly identify the results as "simulations". The authors do label and describe the data as such, however, because of the look of the figure, and in the Discussion, it is very easy to misinterpret as measured experimental data, rather than an in silico sim – perhaps panel shading on a-h would help?
Regarding the simulations, it was not clear what the "imaging" rate was set to be. I am curious why processes where not included in the in silico simulation. I would expect the corrected FOVs to have even better performance than uncorrected, if processes were included, and it would have been nice to explore. Further, I am not quite sure why the authors did not use a standard method to extract the putative ROIs, and compare that with their ground truth – it seems they considered the flaws of standard practice, and biased their sampling of ROIs directly – but I would think using a standard pipeline instead would have been more informative.
Along with this, in the real experiments, different methods were used to pick the ROIs – I am wondering if the authors could comment on the rational for the choices? I say this also, because while one would prefer to record more faithful signals directly, it isn't clear to me that some of the additional "mixing" that would occur between the ROIs would not be removed through source factorization.
The slow imaging rate also leads to higher correlations. Did the authors also collect signals at high frame rates? The basic optical properties would not change at higher rates, but 2-3 Hz is far from typical functional imaging rates. The gains in SNR may be more material at these higher frame rates too. Was it simply a microscope limitation that prohibited faster imaging?
For Figure 7, It is not clear to me why there should be a bias toward information away from the center assuming (semi) random placement of the GRIN- do the authors think that this is real, or simply a measurement artifact – the center of the FOV is nominally furthest from completely healthy tissue, so the slightly lower information reflects network damage? That would also affect the magnitude (but likely not the fundamental interpretation) of the spatial distribution of cells in the modules.
In this paper, the authors describe a novel approach for improving two-photon imaging through thin GRIN-lens endoscopes by using a nanofabricated corrective lens. They validate the approach with in vitro measurements and simulations, and apply the technique to imaging a variety of samples, most notably somatosensory thalamus in awake, behaving mice.
The technique seems easy to apply and could advance the state of the art in several labs. Although I haven't done such experiments myself, I know from other groups that experiments are often sorely limited by their field of view. This paper presents an exciting solution to this problem.
I am enthusiastic about the work, but point out the following issues:
1) With uncorrected GRIN endoscopes, 3D imaging is possible. Do the corrective lenses affect the axial range that can be scanned this way, and if they do, how does this impact the total of neurons that can be recorded with a corrected vs uncorrected endoscope? Zemax simulations would be sufficient to address this.
2) I think the paper could do a better job of motivating the need for thinner endoscopes, particularly in the Introduction. Figure 5 leaves me wondering whether connections would really be disrupted by a larger endoscope. Figure 3—figure supplement 7 makes it clear that larger endoscopes and cannulae require removing a lot of structures, but the paper leaves it to the reader to imagine the consequences. Have there been published accounts of failure rates for different endoscope sizes? Perhaps your own experiences, or studies of e.g. amygdala or hypothalamus?
3) The methods used to segment/analyze calcium imaging data (subsection “Segmentation of simulated time series”) come off as somewhat straw-man. There are a variety of activity-based tools that can separate mixed signals from overlapping cells (PCA/ICA, NMF, CNMF, etc.), and these have been popular for (1P) endoscopic imaging. I wonder if such methods narrow the gap between uncorrected and corrected SNRs and correlations. The statements at the end of paragraph five of the Discussion seem a bit strong if this possibility has not been explored.
4) How chromatic are these corrections? The authors propose simultaneous functional imaging and optogenetic perturbations with these corrective lenses, but this would require the correction be useful across a large wavelength range (perhaps ~900nm and ~1040nm), which needs more evidence. Chromatic issues can easily affect 2P imaging even over the bandwidth of a femtosecond laser.
Reviewer #1:
In this manuscript, the authors demonstrated an extended field-of-view (eFOV), thin (<500μm) microendoscopes by adding aberration correcting microlenses fabricated bt TPL. Performance of the microlens corrected eFOV endoscopes was demonstrated with simulations on synthetic calcium data and in vivo activity imaging. In addition, the eFOV microendoscopes were used to investigate VPM activity correlated with locomotion, whisker movement, and pupil diameter. Cell-specific encoding in VPM during these behaviors are further scrutinized with statistical and machine learning methods. The presentation of the methods and the results are clear. The methods are useful and reasonably practical for increasing the FOV of ultrathin microendoscopes, which is a major limitation for small GRIN lenses. Overall, I recommend publication of the paper in eLife. Listed below are a number of concerns that we hope the authors can address in their revision to further improve the manuscript.
1) For the design and fabrication of the eFOV-microendoscope:
a) The authors presented four types of GRIN lenses (Type I-IV). Four different models of the GRIN lens are shown in subsection “Corrective lens manufacturing and microendoscope assembly”. It will probably help if the authors can explicitly indicate the Types with the model number.
The correspondence between the type and the model number of the GRIN lenses used in the study is now indicated in paragraph two of subsection “Corrective lens manufacturing and microendoscope assembly”.
b) It is claimed that the experimental result is similar to ray-trace simulation as shown in Figure 2. However, there is no comparison to support the claim. The high-order coefficient in the gradient profile in GRIN lens is usually not included in the optical model used for ray-trace. High-order aberration will thus not be reflected in the simulation. It would be important to know the difference between the simulation result and the experimental result. For example, the authors can add experimental data in Figure 2 or add simulation data in Figure 3.
Following the reviewer’s comment, we ran simulations (now shown in new Figure 2—figure supplements 1-2) to quantify the excitation PSF for the four types of endoscopes and to compare these simulation results with the experimental findings displayed in Figure 3. Similarly to what observed in real experiments, in optical simulations we found that the axial dimension of the PSF in lateral portions of the FOV remained smaller and more similar to the axial dimension of the PSF in the center of the FOV in corrected compared to uncorrected endoscopes. Thus, these new optical simulations predicted enlarged FOV in corrected microendoscopes, a result that we experimentally validated (Figure 3). We edited the text in paragraph two of the Results and figures (new Figure 2—figure supplements 1-2) to include these new findings.
It must be noted that the absolute values of the axial PSF were generally smaller in the optical simulations compared to real measurements. This may be due to multiple reasons. First, as suggested by the reviewer, high-order aberrations were not included in the simulations. Second, in simulations, although the intensity of the excitation PSF was small in lateral portion of the FOV (Figure 2), a Gaussian function could still reliably fit the dim intensity distribution and provide a clear quantification of the PSF dimension. In experimental measurements of fluorescence emitted by subresolved beads, the more degraded PSF in the lateral portions of the FOV would result in low efficacy of the excitation beam in stimulating fluorescence, which would result in low SNR fluorescence signals. This would introduce large variability in the fit. Third, small variability in some of the experimental parameters (e.g., the distance between the GRIN back end and the focusing objective) were not reflected in the simulations.
c) For all the experiments, is it performed with the 2-photon polymerized lens with resin or with lens replica using a molded UV-curable adhesive? Is the aberration correction performance different from these two fabrication methods? Which material is the ray-tracing simulation based on?
Experiments and optical characterization were performed using lens replica only. Ray-trace simulation were performed considering the material used in lens replica (i.e., NOA63). We edited the text to make this clearer.
d) What is the yield of the lens fabricated? How is the cured UV-curable adhesive detached from the PDMS mold? During imaging sessions, was there any damage on the lens?
We thank the reviewer for their comments and we inserted the requested information in the text. Specifically:
“The yield for 3D printed lenses and lens replica was ~ 100 %.”
During molding, one drop of UV-curable adhesive was first put on a coverslip, which was then pressed against the mold. One side of the UV-curable adhesive was in contact with the mold, the other side was instead attached to the coverslip. After UV curing, by gently pulling the glass coverslip away the lens made of UV curable adhesive detached easily from the PDMS mold, while remaining firmly attached to the coverslip. This is now more clearly described in subsection “Corrective lens manufacturing and microendoscope assembly”.
We observed no appreciable damage on the lens over imaging sessions (now reported in subsection “Optical characterization of eFOV-microendoscopes”). In considering this, please note that laser beam is not directly focused on the polymer lens, but rather about 100 µm above it (Figure 1).
e) How does the correction microlens affect the focal length of the GRIN lens. In other words, what is the allowable change in working distance before and after the correction?
The simulated focal length in the absence and presence of the corrective microlens for the four different types of endoscopes is now reported in a novel table (new Supplementary file 1-table 2). The difference in focal length between uncorrected and corrected endoscopes is in the range 2-23 µm.
f) There appears to be a much larger field curvature with the corrected GRIN lens. It would be helpful to discuss the impact of this on functional imaging.
A short paragraph describing the impact of the larger field curvature of corrected endoscopes on functional imaging has been added in the Discussion. It reads: “Corrected endoscopes are characterized by a curved FOV. In the case of type II corrected endoscopes, the change in the z coordinate in the focal plane can be up to 75 µm (Figure 3). This z value is smaller for all other corrected endoscope types (Figure 3). The observed field curvature of corrected endoscopes may impact imaging in brain regions characterized by strong axially organized anatomy (e.g., the pyramidal layer of the hippocampus), but would not significantly affect imaging in regions with homogeneous cell density within the z range described above (< 75 µm for type II corrected microendoscopes).”
g) Instead of AO with an active element (e.g., SLM), one could also put a fixed lens with the desired curvature in the place of the SLM. Compared to the microlens approach, one needs to modify the excitation path somewhat, but the lens can be fabricated using standard techniques and may be more accessible to most research labs. It might be informative to the readers to comment on this alternative approach.
Following the reviewer’s comment, we modified the text to discuss this possibility. The new text reads: “A potential alternative to the approach describe in this study would be to place a macroscopic optical element of the desired profile in a plane optically conjugated to the objective back aperture along the optical path. This solution could have the advantage of being manufactured using more standard techniques. However, it would require significant change in the optical set-up, in contrast to the built-in correction method that we describe in the present study. Moreover, this macroscopic optical element would have to be changed according to the type of microendoscope used”.
2) For the data taken with the eFOV-microendoscope, there is ~2x improvement in FOV for type III lens (Figure 3g). However, in Figure 3—figure supplement 5C, the images before and after correction for type III lens is almost the same.
We substituted the images in Figure 3—figure supplement 5C with new ones that are more representative of the results shown in Figure 3G.
3) For the data simulation and analysis:
a) Analysis was shown to correlate whisking, locomotion and cell-specific encoding. Figure 6C and 6D show that locomotion is usually correlated with a higher ΔF/F. Could this be a result of the motion artifact during locomotion from the head-constraint mouse?
We think it is unlikely that the observed increase in ΔF/F is due to motion artifacts during locomotion because:
b) To simulate synthetic calcium data, different models (Gaussian Mixture, lognormal, etc.) are used to obtain the pixel intensity value as described in subsection “Generation of fluorescence time series”. It is mentioned that the model is obtained by fitting the experimental data. What criteria (e.g. comparing likelihood) is used to choose those specific models? What other distribution is tested to choose the best-fit model?
We fitted the experimental data using four alternative models: a normal, a γ, a log-normal distribution, and a Gaussian mixture model. We selected the best-fit model by maximizing the log-likelihood.
In the following tables, the log-likelihood values of the used distributions pixels in the center and lateral portion of the FOV are reported for the reviewer’s inspection (highlighted in green the one we used under each condition):
Central pixels (without ROIs)
Central pixels (with ROIs)
“
c) As mentioned in subsection “Analysis of field distortion and calibration of the pixel size”, the author used a gaussian profile to define axial resolution. It is probably better to fit with a Lorentzian function I(z)=〖(1/(1+〖(z/zR)〗^2 ))〗^2 for 2-photon excitation, where zR is the Rayleigh range of the gaussian focus. Not sure how large the difference will be between these two fits.
Following the reviewer’s comment, we used the Lorentzian function to fit the fluorescence intensity distribution for type II microendoscopes. Axial resolution values computed using the fit with the Lorentzian function were not significantly different compared to those obtained fitting with a Gaussian function (Wilcoxon matched-pairs signed rank test, p = 0,93, n = 132 PSF measurements). We therefore decided to leave the results obtained with the Gaussian function in the manuscript.
d) Figure 4D. It is not clear to me why the ROI number is smaller for the corrected lens than the uncorrected when the peak SNR value is small (e.g., less than 10). Can the authors add some explanations?
We thank the reviewer for raising this point. The smaller number of ROIs when the SNR is low in corrected endoscopes is due to:
1) In the segmentation method we implemented, which is based on the ground truth distribution of the neurons in the simulated sample, at least two pixels belonging to a ground truth neuron are defined as a ROI.
2) In uncorrected endoscopes, the axial PSF largely increases as a function of the radial distance. The enlarged axial PSF in the lateral portions of the FOV augments the probability of sampling voxels belonging to multiple neurons located at different z positions. Once projected in the 2D plane, the contribution of multiple neurons located at different z positions increases the probability of having pixels belonging to ROIs. An increased axial PSF thus leads to an increased number of detected ROIs.
3) Corrected endoscopes have smaller axial PSF compared to uncorrected ones and thus smaller number of detected ROIs.
We added a short text to better explain this result in subsection “Higher SNR and more precise evaluation of pairwise correlation in eFOV-microendoscopes”.
e) The authors claim that "pairwise correlation increased as a function of the radial distance of the pair in uncorrected probes compared to corrected ones." Figure 4M does not seem to show noticeable increase as a function of the radial distance for the uncorrected lens (left panel). Rather, it seems to show a noticeable decrease as a function of the radial distance for the corrected lens (right panel). Therefore, the statement does not seem to be supported by Figure 4M. This is confusing, and some explanation is needed.
We thank the reviewer for raising this point and we apologize with them if the indicated sentence was imprecise and generated confusion. What we meant was that the linear fit of pairwise correlations as a function of radial position for uncorrected endoscopes had a significantly positive slope (Figure 4M, left panel, slope = 0.0002, permutation test p = 0.006), indicating higher pairwise correlations in lateral compared to more central portions of the FOV. For corrected endoscopes, the slope of the linear fit was not significantly different from zero (Figure 4M, right panel, slope = -0.0005, permutation test p = 0.05 for dataset 1; slope = -0.0007, permutation test p = 0.05 for dataset 2). We modified the text to correct the previous imprecision and clarify this point.
4) There is a typo in the first sentence of subsection “Fabrication of eFOV-microendoscopes”. I believe it is (Figure 3—figure supplement 1C) instead of (Figure 1C).
Thank you for spotting this error. It has been fixed.
Reviewer #2:
In this paper, the authors describe the creation and use of microfabricated corrector lenses for improving GRIN lens imaging. GRIN lenses are one of a few technologies (and the only one that is "mature") that allow for high resolution optical imaging in brains at depths beyond ~1.5 mm. These correction lenses a) increase the usable field-of-view (FOV) of a given GRIN lens, and b) improve the fidelity of signals extracted from regions-of-interest (ROIs), by decreasing the point spread function (PSF) volume. Both are important for high fidelity recording of neuronal populations. The paper is well written, and I find their demonstrations compelling. I think the work is of broad interest, and suitable for publication in eLife, after addressing a few concerns and questions.
My comments below are mainly critical, but I note that I think the paper overall is very good, and thorough, and I have high enthusiasm.
Given that the surface profiles begin their optimization from an analytical expression, It would be nice if the authors could comment on the likely maximal possible correction possible – that is on a GRIN lens with a GRIN relay of significant pitch, do they think microlenses would be similarly manufacturable or effective?
We thank the reviewer for this interesting question. It is expected that optical aberrations of GRIN-rod lenses become more pronounced as the GRIN-rod length increases. This raises the possibility that the needed correcting lens profile could significantly differ from our current initial guess (the Schmidt corrector plate profile). To properly answer the reviewer’s question, there are thus a number of issues that need to be evaluated altogether, namely: a) up to what length of the GRIN-rod, for a given NA, the Schmidt corrector plate profile is an effective starting guess for the lens optimization process; b) once such an optimized correcting profile is found, to what extent the introduced correction is able to increase the effective FOV; c) how accurately the complex correcting lens profile could be reproduced by 3D micro-printing. Because our present approach satisfactory works for the GRIN-rod lengths considered in our study (1.1 mm – 4.07 mm), and because of the amount of work needed to address the manifold and interrelated aspects of the asked question, we deemed it to be the subject of a whole new study. As consequence, we currently cannot provide a final answer to this question, but we aim at addressing the matter in the continuation of our work. Please note that in the paper we do not make any claim that the method is applicable to GRIN lenses of any length, and we made sure in revision that this limitation in scope is acknowledged.
In the simulated imaging data, especially the figure, and SNR conclusions drawn from that, I think it is important to more clearly identify the results as "simulations". The authors do label and describe the data as such, however, because of the look of the figure, and in the Discussion, it is very easy to misinterpret as measured experimental data, rather than an in silico sim – perhaps panel shading on a-h would help?
We thank the reviewer for their suggestion. We followed the indication of the reviewer and modified Figure 4 to make clear, using boxes, which data come from simulations and which from real experiments.
Regarding the simulations, it was not clear what the "imaging" rate was set to be.
In the simulations the “imaging” rate was set to 5 Hz. This is indicated in subsection “Generation of fluorescence time series”.
I am curious why processes where not included in the in silico simulation. I would expect the corrected FOVs to have even better performance than uncorrected, if processes were included, and it would have been nice to explore.
We did not address the impact of including processes in the simulations, because the axial PSF of the GRIN microendoscopes was in the range 8-10 µm and the labeling that we achieved in our preparation was rather dense. Under these experimental conditions, resolving individual neuronal processes was complicated and we limited our goal to image fluorescent signals from neuronal cell bodies. We agree with the reviewer this is an interesting point, which can be addressed in future investigations.
Further, I am not quite sure why the authors did not use a standard method to extract the putative ROIs, and compare that with their ground truth – it seems they considered the flaws of standard practice, and biased their sampling of ROIs directly – but I would think using a standard pipeline instead would have been more informative.
Following the reviewer’s comment, we first compared the quality of the manual segmentation described in this study with that of a standard automated algorithm (e.g., CaImAn, (Giovannucci et al., 2019) by computing precision, recall, and F1 score in simulated data (new Figure 4—figure supplement 1). We found that the automated method was characterized by recall values which were, on average across SNR values, < 0.4 (new Figure 4—figure supplement 1A), leading to the detection of only a minority of the total number of neurons. In contrast, the manual method led to higher recall across SNR threshold values. Notably, both manual and automated methods yielded larger recall in corrected endoscopes compared to uncorrected ones. Moreover, for low SNR threshold values the automated segmentation had precision values < 0.8 in both uncorrected and corrected endoscopes (new Figure 4—figure supplement 1B), leading to identification of ROIs which did not correspond to cells in the ground truth. In contrast, the manual segmentation method had much larger values of precision across SNR threshold levels. Overall, F1 scores were higher for the manual segmentation method compared to the automated one for both uncorrected and corrected endoscopes (new Figure 4—figure supplement 1C). Because of these results showing better performance of the manual segmentations in simulated data (an observation now included in the revised manuscript), we took the decision to present in the main figures of the paper results based on manual segmentation.
In the revised manuscript, we extended the comparison between the manual and automated segmentation methods to real data. We observed that in uncorrected endoscopes CaImAn identified smaller number of ROIs compared to manual segmentation (new Figure 4—figure supplement 1D). In contrast, the number of ROIs identified with CaImAn and the manual method in t-series acquired with the corrected endoscope were not significantly different (new Figure 4—figure supplement 1E). One potential explanation of this finding is that the automated segmentation method more efficiently segments ROIs with high SNR compared to the manual one. Since aberration correction significantly increases SNR of fluorescent signals, the automated segmentation performed as the manual segmentation method in corrected endoscopes.
Overall, we believe that, in the context of our study, the use of the manual segmentation method rather than an automated segmentation algorithm is still preferable, because it allowed a more precise evaluation of ROIs in simulated data and fairer comparison of corrected and uncorrected endoscopes in real data. Most importantly, improvements introduced by corrected endoscopes could be observed with both the manual and the automated segmentation methods (see response to the next point). We believe that the comparisons suggested by reviewer are of interest and we feel that the new results in Figure 4—figure supplement 1 improve our manuscript. We are grateful to the reviewer for this suggestion.
Along with this, in the real experiments, different methods were used to pick the ROIs – I am wondering if the authors could comment on the rational for the choices? I say this also, because while one would prefer to record more faithful signals directly, it isn't clear to me that some of the additional "mixing" that would occur between the ROIs would not be removed through source factorization.
In all experiments presented in the main paper, we used our customized manual pipeline for the analysis of fluorescence t-series. In Figure 3—figure supplement 6 we used a different method, the PCA/ICA algorithm described in Mukamel et al., Neuron 2009, to segment neurons and extract calcium signals. This was originally done to show that data recorded with corrected endoscopes could be analyzed with standard automated methods. Following the comments of this reviewer and that of reviewer #3 asking for a systematic evaluation of how much the improvements introduced by aberration correction depend on the analysis pipeline, we removed the PCA/ICA analysis in Figure 3—figure supplement 6 and we systematically evaluated the effect of aberration correction on the output of the analysis in the simulated and experimental data shown in Figure 4 using both the manual segmentation method and an automated segmentation method (e.g., CaImAn).
In simulated data, we found that with CaImAn the number of ROIs segmented in corrected endoscopes was consistently higher than in uncorrected endoscopes across SNR thresholds (new Figure 4—figure supplement 2A), similarly to what observed with manual segmentation (Figure 4D). Using CaImAn, SNR values of fluorescence events were significantly higher in corrected compared to uncorrected endoscopes (new Figure 4—figure supplement 2B), similarly to what observed with manual segmentation (Figure 4E). Moreover, the linear fit of pairwise correlations as a function of radial position for uncorrected endoscopes had a significantly positive slope (new Figure 4—figure supplement 2C left panel), indicating higher pairwise correlations in lateral compared to more central portions of the FOV. For corrected endoscopes, the slope of the linear fit was not significantly different from zero (new Figure 4—figure supplement 2C right panel). This result is also in line with what we previously observed with manual segmentation (Figure 4G). The description of these novel results is added in the text.
In experimental data, we found that SNR values of fluorescence events tended to be higher in corrected compared to uncorrected endoscopes (new Figure 4—figure supplement 2D), a trend that was in line with what observed with manual segmentation (Figure 4L). The slope of the linear fit of pairwise correlations as a function of radial position for uncorrected endoscopes was significantly positive (new Figure 4—figure supplement 2E), indicating higher pairwise correlations in lateral compared to more central portions of the FOV. For corrected endoscopes, the slope of the linear fit was not significantly different from zero (new Figure 4—figure supplement 2E). Both results are in line with what previously observed with manual segmentation (Figure 4M). The description of these novel results is added in the text.
Overall, the results of the systematic comparison between the manual and automated segmentation methods show that improvements introduced by aberration correction in endoscopes could be observed with both the manual and the automated segmentation methods. We thank the reviewer for raising this point.
The slow imaging rate also leads to higher correlations. Did the authors also collect signals at high frame rates? The basic optical properties would not change at higher rates, but 2-3 Hz is far from typical functional imaging rates. The gains in SNR may be more material at these higher frame rates too. Was it simply a microscope limitation that prohibited faster imaging?
We did not collect fluorescent signals at frame rates higher than 4 Hz. This was because we aimed at imaging the largest possible FOV and our experimental set up was equipped with regular galvanometric mirrors and not with resonant mirrors. We added a sentence in the text to highlight this limitation of our study.
For Figure 7, It is not clear to me why there should be a bias toward information away from the center assuming (semi) random placement of the GRIN- do the authors think that this is real, or simply a measurement artifact – the center of the FOV is nominally furthest from completely healthy tissue, so the slightly lower information reflects network damage? That would also affect the magnitude (but likely not the fundamental interpretation) of the spatial distribution of cells in the modules.
At the single cell level, information was evenly distributed across the FOV (Figure 7C). We believe the reviewer’s comment refers to data shown in Figure 7E, which contains data pooled from a population of neurons. In Figure 7E, we observed an increase in the information carried by the population of neurons when we incrementally considered larger populations by adding ROIs at larger radial distances. This does not imply that ROIs away from the center carry higher information, but only that their information sums to the information of more central ROIs in a way that is not purely redundant. We apologize if the description of Figure 7E was confusing, and we edited the text to make it clearer.
Reviewer #3:
In this paper, the authors describe a novel approach for improving two-photon imaging through thin GRIN-lens endoscopes by using a nanofabricated corrective lens. They validate the approach with in vitro measurements and simulations, and apply the technique to imaging a variety of samples, most notably somatosensory thalamus in awake, behaving mice.
The technique seems easy to apply and could advance the state of the art in several labs. Although I haven't done such experiments myself, I know from other groups that experiments are often sorely limited by their field of view. This paper presents an exciting solution to this problem.
I am enthusiastic about the work, but point out the following issues:
1) With uncorrected GRIN endoscopes, 3D imaging is possible. Do the corrective lenses affect the axial range that can be scanned this way, and if they do, how does this impact the total of neurons that can be recorded with a corrected vs uncorrected endoscope? Zemax simulations would be sufficient to address this.
The optical simulations we did to design the corrective lenses were performed maximizing aberration correction only in the focal plane of the endoscope. Following the reviewer’s comment, we explored the effect of aberration correction outside the focal plane using new Zemax simulations. In corrected endoscopes, we found that the Strehl ratio was > 0.8 (Maréchal criterion) in a larger volume compared to uncorrected endoscopes (new Figure 2—figure supplement 3). The range of volume with high Strehl ratio in corrected endoscopes was 1.7-3.7 larger than the volume with high Strehl ratio in uncorrected probes, leading to similarly larger number of neurons being imaged in the enlarged volume. This is now described on lines (paragraph three subsection “Optical simulation of eFOV-microendoscopes”).
2) I think the paper could do a better job of motivating the need for thinner endoscopes, particularly in the Introduction. Figure 5 leaves me wondering whether connections would really be disrupted by a larger endoscope. Figure 3—figure supplement 7 makes it clear that larger endoscopes and cannulae require removing a lot of structures, but the paper leaves it to the reader to imagine the consequences. Have there been published accounts of failure rates for different endoscope sizes? Perhaps your own experiences, or studies of e.g. amygdala or hypothalamus?
Following the reviewer’s comment, we modified the Introduction to more clearly motivate the need of thin endoscopic probes for deep brain functional imaging in paragraph two of the Introduction. We are not aware of a systematic description on the impact of different endoscope sizes on the functionality of the targeted brain region, but we have cited relevant literature showing that surgical lesion of reciprocal connectivity alters thalamocortical dynamics.
3) The methods used to segment/analyze calcium imaging data (subsection “Segmentation of simulated time series”) come off as somewhat straw-man. There are a variety of activity-based tools that can separate mixed signals from overlapping cells (PCA/ICA, NMF, CNMF, etc.), and these have been popular for (1P) endoscopic imaging. I wonder if such methods narrow the gap between uncorrected and corrected SNRs and correlations. The statements at the end of paragraph five of the Discussion seem a bit strong if this possibility has not been explored.
Following the reviewer’s comment we evaluated the effect of aberration correction on the output of the analysis in the simulated and experimental data shown in Figure 4 using both the manual segmentation method and an automated segmentation method (e.g., CaImAn).
In simulated data, we found that with CaImAn the number of ROIs segmented in corrected endoscopes was consistently higher than in uncorrected endoscopes across SNR thresholds (new Figure 4—figure supplement 2A), similarly to what observed with manual segmentation (Figure 4D). Using CaImAn, SNR values of fluorescence events were significantly higher in corrected compared to uncorrected endoscopes (new Figure 4—figure supplement 2B), similarly to what observed with manual segmentation (Figure 4E). Moreover, the linear fit of pairwise correlations as a function of radial position for uncorrected endoscopes had a significantly positive slope (new Figure 4—figure supplement 2C left panel), indicating higher pairwise correlations in lateral compared to more central portions of the FOV. For corrected endoscopes, the slope of the linear fit was not significantly different from zero (new Figure 4—figure supplement 2C right panel). This result is also in line with what previously observed with manual segmentation (Figure 4G). The description of these novel results is added in subsection “Higher SNR and more precise evaluation of pairwise correlation in eFOV-microendoscopes”.
In experimental data, we found that SNR values of fluorescence events tended to be higher in corrected compared to uncorrected endoscopes (new Figure 4—figure supplement 2D), a trend that was in line with what observed with manual segmentation (Figure 4L). The slope of the linear fit of pairwise correlations as a function of radial position for uncorrected endoscopes was significantly positive (new Figure 4—figure supplement 2E), indicating higher pairwise correlations in lateral compared to more central portions of the FOV. For corrected endoscopes, the slope of the linear fit was not significantly different from zero (new Figure 4—figure supplement 2E). Both results are in line with what previously observed with manual segmentation (Figure 4M).
The description of these novel results is added in subsection “Higher SNR and more precise evaluation of pairwise correlation in eFOV-microendoscopes”.
Overall, the results of the systematic comparison between the manual and automated segmentation methods show that improvement introduced by aberration correction in endoscopes could be observed with both the manual and automated segmentation methods. We thank the reviewer for raising this point.
4) How chromatic are these corrections? The authors propose simultaneous functional imaging and optogenetic perturbations with these corrective lenses (Discussion final paragraph), but this would require the correction be useful across a large wavelength range (perhaps ~900nm and ~1040nm), which needs more evidence. Chromatic issues can easily affect 2P imaging even over the bandwidth of a femtosecond laser.
Thanks for this comment. All optical simulations described in the first submission were performed at a fixed wavelength (λ = 920 nm). Following the reviewer’s request, we explored the effect of changing wavelength on the Strehl ratio using new Zemax simulations. We found that the Strehl ratio remains > 0.8 within ± 15 nm from λ = 920 nm (new Figure 2—figure supplement 4), which covers the limited bandwidth of our femtosecond laser but not that needed for all-optical imaging and manipulation experiments. To perform this latter type of experiments, the SLM that is used in the stimulation pathway should also be used to correct for the z-defocus with an appropriate lens function. Alternatively, future microfabrication work will be needed to design more complex corrective lens design, which efficiently compensates chromatic aberrations in the 900-1100 nm wavelength range. These possibilities are now more clearly discussed (Discussion paragraphs eight and nine).