Seeing Natural Images through the Eye of a Fly with Remote Focusing Two-Photon Microscopy.

Visual systems of many animals, including the fruit fly Drosophila, represent the surrounding space as 2D maps, formed by populations of neurons. Advanced genetic tools make the fly visual system especially well accessible. However, in typical in vivo preparations for two-photon calcium imaging, relatively few neurons can be recorded at the same time. Here, we present an extension to a conventional two-photon microscope, based on remote focusing, which enables real-time rotation of the imaging plane, and thus flexible alignment to cellular structures, without resolution or speed trade-off. We simultaneously record from over 100 neighboring cells spanning the 2D retinotopic map. We characterize its representation of moving natural images, which we find is comparable to noise predictions. Our method increases throughput 10-fold and allows us to visualize a significant fraction of the fly's visual field. Furthermore, our system can be applied in general for a more flexible investigation of neural circuits.

Introduction

Vision is a crucial sense for many animals, including humans. Through vision, we gather abundant information about our surroundings, which is processed in the visual system to extract relevant features that then guide behavioral choices. In the initial stages of processing, visual information is represented as a set of 2D maps of the surrounding space (Chklovskii and Koulakov, 2004, Land and Fernald, 1992). In the fruit fly Drosophila melanogaster, the visual system collects information through 750 ommatidia (Ready et al., 1976), each containing a lens and eight photoreceptors. Correspondingly, the subsequent neuropils, lamina, medulla, lobula, and lobula plate are each composed of 750 columns. In the lamina, photoreceptor inputs from neighboring ommatidia converge such that one lamina column receives information from one point in visual space (“neural superposition principle,” Braitenberg, 1967, Kirschfeld, 1967). Neighboring columns process information from neighboring points in space (Land, 1997). Together, they thus retinotopically map the surrounding visual scene. Accordingly, many cell types, which code for a specific visual feature like local contrast or motion, come in 750 unicolumnar copies (for recent reviews, see, e.g., Mauss et al., 2017, Song and Lee, 2018).

Likewise, the vertebrate retina is composed of several layers of neurons that form retinotopic maps, each of which processes certain visual features (for a comparative review, see, e.g., Borst and Helmstaedter, 2015, Sanes and Zipursky, 2010). Importantly, due to the retina's anatomical separation from the rest of the brain, it can be mounted on a flat surface and the retinotopic array of neurons can be recorded from by means of microelectrode arrays or two-photon calcium imaging (Baden et al., 2016, Denk and Detwiler, 1999, Meister et al., 1994, Pillow et al., 2008, Segev et al., 2004). This ease of access has probably made the retina one of the best-understood neural circuits (for reviews, see Demb and Singer, 2015, Diamond, 2017, Dowling, 2012).

The fruit fly poses many advantages for the study of neural circuits. Morphologically and genetically defined cell types, stereotyped anatomy, and refined genetic tools for targeted expression, recording, and manipulation, together with two-photon imaging, resulted in an unparalleled understanding of the Drosophila visual system (for reviews, see Aptekar and Frye, 2013, Mauss et al., 2017, Song and Lee, 2018). However, due to the differences in their anatomy, the retinotopic maps cannot be recorded from as in the mammalian retina. The retinotopic plane in the Drosophila visual system, as typically mounted for in vivo calcium imaging and visual stimulation, is an arbitrarily rotated plane with respect to the imaging system (Figures 1A and 1B, Reiff et al., 2010, Seelig et al., 2010). Its precise orientation depends on the specifics of the mounting and setup arrangement. With a scanning two-photon microscope, only the cells that lie in the focal plane can be recorded from at fast scan speeds (Figure 1C). Thus, functional analysis of the visual system is restricted to few cells at a time, which limits the throughput and analysis of more complex stimuli, for which recording from a larger population would be required.

Figure 1

Imaging the Drosophila Visual System

(A) Setup for in vivo calcium imaging of the Drosophila visual system during visual stimulation. The fly is fixed in a holder while the brain is viewed from the dorso-posterior side. Visual stimuli are shown on a cylindrical screen surrounding the fly. Schematic not to scale.

(B) The visual system of Drosophila consists of retinotopically arranged layers of neuropil. The horizontal imaging plane (purple) sections the visual system in a particular orientation. An oblique imaging plane (turquoise) can be aligned to the retinotopic plane of the visual system. The axon terminals of the cell type L2 lie in layer 2 of the medulla. For simplicity, the chiasm between lamina and medulla is omitted in this schematic.

(C) Horizontal plane of L2 expressing GCaMP6f.

(D) Oblique, retinotopic plane of axon terminals of L2 in the medulla layer 2 expressing GCaMP6f.

Recent advances in imaging technologies are opening doors to a more flexible investigation of neural structures (for reviews. see Ji et al., 2016, Ronzitti et al., 2018). For example, random-access two-photon microscopy using acousto-optic modulators enables fast access to sparse cell locations (Katona et al., 2012), and fast z-scanning is possible among others with electrically tunable lenses or remote focusing (Botcherby et al., 2008, Grewe et al., 2011). Employing the remote-focusing principle, the focal spot of the laser is moved fast along the optical axis by means of moving a third, remote mirror (Botcherby et al., 2008). When certain criteria are met, imaging can be aberration free in a large volume (Botcherby et al., 2007, Botcherby et al., 2012). Remote focusing has primarily been applied to skip to several horizontal planes (parallel to the focal plane) with fast imaging rates (Rupprecht et al., 2016, Sofroniew et al., 2016). In principle, however, it should enable other scanning trajectories as well. Here, we present a near-aberration-free remote-focusing module, which can be added onto conventional two-photon microscopes and which can be controlled by an extension to the open source software ScanImage 5.1 (Pologruto et al., 2003). Our software extension facilitates online rotation of the imaging plane in 3D in any arbitrary direction. The plane can thus be aligned with cellular structures of interest during the imaging session, and oblique planes can be imaged without a trade-off in resolution or speed.

We apply this method to simultaneously image over 100 axon terminals in the retinotopic plane of the unicolumnar cell types L2 (Figure 1D) and Mi1 and thus reconstitute the 2D representation of the visual scene as represented by these cell types. Both cell types play a crucial part in the circuit for motion vision (Ammer et al., 2015, Joesch et al., 2010, Strother et al., 2017, Tuthill et al., 2013). L2 receives direct input from photoreceptors (Rivera-Alba et al., 2011) and codes for local light decrements when probed with artificial stimuli (Clark et al., 2011, Freifeld et al., 2013, Reiff et al., 2010). Mi1 codes for local light increments (Arenz et al., 2017, Behnia et al., 2014) and provides output to direction-selective T4 cells (Strother et al., 2017, Takemura et al., 2017). Here, we characterize the 2D population representation of natural images by these cell types. Based on their functional properties derived from experiments with artificial stimuli we expect that they code for local luminance changes of natural scenes.

Results

Imaging Oblique Planes with Remote Focusing Two-Photon Microscopy

We first designed a remote-focusing module that is easily integrated into a typical two-photon microscope and provides near-homogenous resolution throughout a large volume (Figure 2A). The remote-focusing principle is based on creating a beam with a convergent or divergent waveform, which, upon passing through a lens, will focus before or beyond the lens' focal plane, respectively (Figure 2B, ; Botcherby et al., 2008). To create such a beam and to easily modulate the beam vergence, we introduced a second objective lens (remote objective lens), and a mirror (Z-mirror, Figure 2A inset, Figure 2B, left), moveable with a piezo motor. The collimated laser beam passes through the remote lens, is reflected by the z-mirror, and is projected back through the lens. Depending on the position of the z-mirror relative to the focal plane of the lens, the resulting waveform is convergent or divergent (Figure 2B). To image the back aperture of the remote lens onto the back focal plane of the imaging objective lens, we used two telescopes (Figures 2A and 2C), with the intermediate conjugate plane at the lateral scan unit. The magnifications of the telescopes were matched such that the overall magnification between the remote and imaging lens was close to the ratio between the refractive indices of water and air, n2/n1 (for details, see Transparent Methods). This way, the aberrations deliberately introduced by the remote lens are to a large extent canceled out by the imaging objective lens (Botcherby et al., 2008). As a result, the optical resolution is near-diffraction-limited throughout the imaging volume, and the relationship between mirror position and position of the laser focal spot in z is linear (Botcherby et al., 2012).

Figure 2

Remote Focusing Two-Photon Microscope

(A) The remote-focusing module in a standard two-photon microscope. A polarization-based beam splitter (PBS) and λ/4 plate guide the beam to the axial scan unit, which is composed of the remote objective lens (ROL) and the piezo-motorized z-mirror. Two lenses (f1 and f2, focal lengths of 125 and 40 mm, respectively) in 4f configuration relay the beam from the back aperture of the ROL to the lateral scan unit. Scan and tube lens (f3 and f4, focal lengths of 50 and 200 mm, respectively) then relay the beam to the imaging objective lens (IOL). Inset: Photograph of the axial scan unit.

(B) Schematic of the beam waveform at the ROL and IOL. A collimated beam entering the ROL is reflected by the z-mirror. Depending on the mirror position relative to the nominal focal plane, the exiting beam at the back focal plane of the ROL will have non-zero vergence. The beam relayed to the IOL results in a focal spot outside the nominal focal plane of the IOL.

(C) Photograph of the remote-focusing module.

(D) Point spread function (PSF) of 0.1 μm-diameter fluorescent beads at five different positions in z in their lateral (x-y) profile. Top: Bead images, average of three measurements. Bottom: 1D profiles along the x axis.

(E) As in (D) for the axial (x-z) profile of the PSF. Top: Bead images, average of three measurements. Bottom: 1D profiles along a tilted axis corresponding to the widest point of the PSF, determined by fitting a 2D Gaussian to the bead images. 1D Gaussian fit along the tilted axis in dashed black lines.

(F) Calcium imaging of the retinotopic plane of Mi1 axon terminals in layer 10 of the medulla. Top left: Calcium baseline level without a stimulus. Bottom left: A letter is displayed on the stimulus screen centered on the receptive fields of the cells in view. Right: Fluorescence increase in response to different letters. Scale bar 20 µm.

To quantify the optical resolution across the imaging volume in our setup, we measured the point spread function of fluorescent beads at different positions in a volume of 100 μm3. The lateral resolution, defined as the full width at half maximum of the bead point spread function, lies below 0.75 μm at all positions in the volume (Figures 2D, 2E, S1, and S2), close to the theoretical estimate (for details see Transparent Methods). The axial resolution is on average 3.6 μm (ranging from 3 to 5 μm at all positions in the volume; Figures 2D, 2E, S1, and S2). The mirror movement causes near-linear displacement of the focal spot in the axial direction (Figure S1I) across 90 μm.

To utilize the remote-focusing module for imaging of oblique planes, we extended the functionality of the open source software ScanImage 5.1 (Pologruto et al., 2003) to control the movement of the motorized z-mirror. In ScanImage, the scanned imaging plane is represented as a sequence of 2D vectors, each dimension corresponding to the movement of one mirror, the lateral scan mirrors (x- and y-mirrors). We extended this representation to a third dimension, reflecting the movement of the z-mirror. Applying geometric operations, the imaging plane can then be rotated and translated arbitrarily in 3D, including simultaneous rotations about several axes. Finally, additional user controls in the ScanImage GUI enable plane rotation in real time, such that the user can align the plane to cellular structures during imaging.

A custom-build piezo motor was used to move the z-mirror (for details see Transparent Methods). Together with two galvanometer motors moving the mirrors in the lateral scan unit, this enabled imaging of arbitrarily rotated, oblique planes of size 90 × 90 μm at frame rates of up to 15 Hz, or smaller planes (50 × 50 μm) at frame rates up to 20 Hz. These frame rates are typical for recordings of the visual system in Drosophila with fluorescent indicators.

To demonstrate the applicability of our method, we performed calcium imaging experiments from the medulla interneuron Mi1. We expressed the calcium indicator GCaMP6f in Mi1 and aligned the oblique imaging plane to the retinotopic map spanned by the cells' axon terminals, viewing the 2D array of axon terminals in the imaging plane (Figure 2F, top left). We then presented different letters in the fly's field of view, each for 1 s (Figure 2F, bottom left). Cells with a receptive field within the pixels of the letter responded with an increase in calcium levels (Figure 2F, right). As a result, the activity map of Mi1 clearly reflects the respective letter shown to the fly.

With our method, we can simultaneously image up to 100 axon terminals in a single field of view, which is a substantial fraction of the fly's visual field. Moreover, this constitutes a 10-fold increase in throughput compared with imaging in the horizontal plane, without compromising spatial resolution or scan speed.

Linear Receptive Field Properties of L2

We next utilized this technique to map the 2D arrangement of receptive fields of the lamina cell type L2. This will subsequently allow us to assign each cell to the location in space it is centered at, and to compare its linear response properties to its responses to natural scene stimulation. To this end, we expressed the calcium sensor GCaMP6f in L2 and aligned the imaging plane with the retinotopic plane spanned by L2 axon terminals (Figure 3A). Each cell in the imaging field of view was easily distinguishable due to its compact axon terminal. We then showed a white noise stimulus (for details see Transparent Methods and Arenz et al., 2017) to the fly while recording fluorescence signals. From each terminal, we extracted the calcium signal and reverse-correlated it to the stimulus, thus obtaining its spatiotemporal receptive field (Dayan and Abbott, 2005). The receptive field location of a cell, the point in visual space it responds to, is then given by the coordinates of the maximum absolute correlation (Figure 3B). The hexagonal and retinotopic coverage of visual space by L2 is visualized in Figures 3C–3E and reflects the ommatidial layout of the Drosophila eye (Buchner, 1971). As expected from previous studies that probed L2 response properties with flashes, gratings, and noise (Drews et al., 2020, Freifeld et al., 2013, Reiff et al., 2010), L2 responded to light decreases with an increase in calcium levels, as is shown by the negative receptive field center (Figure 3F, N = 12 flies, n = 711 cells). In addition, the spatial receptive field showed an anisotropic, antagonistic surround (Figures 3F and S4), which has been suggested by Freifeld et al. (2013). In the temporal domain, L2 showed band-pass characteristics (Figure 3G). The receptive field size was on average 1.91 ± 0.24° (σ of a Gaussian fit, corresponding to 4.51 ± 0.57° full width at half maximum), and we did not observe systematic differences across the eye.

Figure 3

Receptive Field Analysis of L2 with White Noise

(A) Mean image of a recording of L2 axon terminals in layer 2 of the medulla expressing GCaMP6f. Approximate orientation in the fly brain as indicated with dorsoventral (D-V), anterior-posterior (A-P), and medial-lateral (M-L) axes.

(B) Reverse correlation of the white noise stimulus with the calcium signal of an example cell. The heatmap shows a cross section of the receptive field at the time point with the absolute peak amplitude. Red corresponds to a positive correlation with the stimulus luminance, and blue corresponds to a negative correlation (a.u.). Black traces show 1D cross sections.

(C) Selected regions of interest encompassing single axon terminals colored by their receptive field location in elevation.

(D) Axon terminals colored by their receptive field location in azimuth.

(E) Receptive field locations (black dots) with receptive field size (sigma of Gaussian fit, gray shaded areas) of all cells in the example recording also shown in (A, C, and D).

(F) Average spatial receptive field of all cells (N = 12 flies, n = 711 cells). The profile in azimuth (black line) was fitted with a Gaussian (dashed yellow line).

(G) 1D and 2D temporal profiles of the receptive field.

L2 Responses to Natural Scenes Encode Luminance

We next asked how a population of L2 cells responds to natural images, and whether local image luminance can account for its responses. To this end, we showed natural images to the fly while recording from the retinotopic plane of L2 axon terminals. The stimulus set consisted of 50 images from the van Hateren Image Database (van Hateren and van der Schaaf, 1998), featuring a variety of natural environments. Each image was displayed for 1.5 s, following a gray screen. In addition, an ON and an OFF flicker stimulus, a completely white or black image, respectively, were included in the stimulus set. Before the natural image stimulation, receptive field coordinates of the cells were determined with white noise analysis (Figure 4A). The calcium signals from all cells were baseline subtracted and normalized (for details see Transparent Methods). Responses from three trials were averaged. The population response of L2 cells in an individual recording shows that L2 cells respond to local image luminance with a calcium increase in dark areas and decrease in bright areas of the image (Figure 4B).

Figure 4

L2 Responses to Natural Scenes

(A) Example image with the receptive field location of all cells in an example recording (red circles).

(B) Top left: Example response of a cell to ON and OFF flicker stimuli and an image. Average over three trials. Shaded areas indicate standard deviation. The gray shaded box indicates the time during which the image was shown. Remainder of figure: Example responses of all cells in an example recording (same example as in A). The background shows the image stimulus, and individual plots are positioned on the corresponding receptive field locations of each cell.

(C) Schematic of the model. Stimulus images were first blurred (filtered with a 2D Gaussian kernel, σ = 1.91°). The luminance values of the filtered images at the receptive field coordinates of each cell then constituted the prediction. Prediction values were scaled for each cell to fit the normalized fluorescence values of the data.

(D) Response predictions of the example recording in (A) and (B).

(E) Left: Peak response of all cells in the example recording. The peak response was defined as the largest absolute response during the time the image was shown. Right: Sustained responses of all cells in the example recording. The sustained response was defined as the average response over the last second during which the image was shown.

(F) Sustained responses of an example cell (cell #13) to all 50 images, plotted against the model predictions. Error bars denote the lowest and highest response out of three trials (range). Dashed yellow line represents the linear fit.

(G) Peak responses in dependency of the response prediction (N = 8 flies, n = 290 cells). Hexagon color represents the number of data points in each hexagon. Green line represents the average; shaded area indicates 95% confidence interval (over binned data). Dashed black line represents linear fit.

(H) Sustained responses as for peak responses in (G).

(I) Correlation between trials (black) and between prediction and data (yellow), expressed as the Pearson's correlation coefficient r. Error bars denote 95% confidence intervals.

To quantitatively compare the cellular response to the local luminance of the image at its receptive field center, we blurred the images with a 2D Gaussian, the standard deviation and sign of the filter given by the receptive field size from the white noise analysis (Figure 4C). This simple model yields a prediction of the cells' responses should they solely respond according to local image luminance (Figure 4D). We evaluated two response features: the peak response was defined as the largest absolute response during the time the image was shown (Figures 4B and 4E, left) and the sustained response was defined as the average response over the last second during which the image was shown (Figures 4B and 4E, right). We then tested whether the quantified responses correlate with model predictions. We found a significant linear dependence of L2 responses to local image luminance (Figures 4F–4H, Bonferroni-corrected Student's t test: p < 0.05 for 286/290 cells). Thus, L2 encodes brightness increments with a calcium decrease and brightness decrements with a calcium increase.

L2 codes for local luminance, yet its responses might in addition depend on other features of the natural scene. Thus, we next asked whether the luminance-based model can account for all response variance. To this end, we first quantified the reliability of the responses as the Pearson's correlation coefficient r between trials (Figure 4I, for details see Transparent Methods). The reliability determines an upper bound for the explainable variance (expressed in r2) by the model; the remaining variance derives from cellular and experimental noise. The model's correlation with the data was comparable to the response reliability, with only a small difference for peak responses (Figure 4I, for statistics see Table S1). Thus, the linear model can explain most of the predictable variance of the responses. We conclude that the retinotopic map of L2 cells, when probed with static natural images, largely represents their local luminance.

L2 and Mi1 Responses to Moving Natural Scenes

L2 represents the major lamina input to the OFF pathway of motion vision (Joesch et al., 2010). Thus, its responses to moving stimuli are particularly relevant to the downstream circuitry. Therefore, we measured the dynamic representation of moving images by the 2D array of L2 axon terminals. We used 10 panoramic images (Figure 5A, Brinkworth and O'Carroll, 2009) and moved them horizontally across the screen at constant speed. Before presenting the moving images, we mapped the receptive fields of individual cells using white noise analysis as done previously. Figure 5B displays example responses of L2 to a selection of moving images (black traces). Figure 5D shows all cells' responses of an example recording at a particular time point during the recording together with the image section viewed at that time point (see also Video S1). To quantify the response properties, we generated response predictions based on three models of increasing complexity (Figure 5F). The first model reports the blurred image luminance at the receptive field center of a cell at a given time point (Luminance model, pink). This model corresponds to the model we used to predict the static natural scene responses. The second model's predictions were generated by filtering the image sequences with the 3D spatiotemporal receptive field (RF model, brown). Compared with the first model, it adds the cells' temporal properties as well as the spatial surround. The third model adds a static nonlinearity to the RF model (L-NL model, blue; for details see Transparent Methods). Exemplary model predictions are shown for the cells in Figure 5B. The correlation between the RF model and the data is shown in Figure 5G. To obtain the static nonlinearity for the L-NL model, we fitted a logistic function to the input-output curve of the RF model and the data (Figures 5G and 5H). The luminance model could predict the data to a certain degree, whereas the RF and L-NL models fared significantly better than the first (Figure 5I, p = 1.49 × 1017 and p = 1.19 × 1017, respectively). We compared the performance of each model with the reliability of the data (determined as for the static natural scenes as the correlation between different trials, as a proxy for the maximally predictable variance). Neither of the models was able to predict the data fully. Thus, although the representation of moving natural images by L2 to a large part reflects its linear response properties, a small part of the response remains unexplained.

Figure 5

L2 and Mi1 Responses to Moving Natural Scenes

(A) Illustration of the stimulus. 360° panoramic images moved to the left for 7 s, followed by motion to the right for 7 s.

(B) Example responses of L2 cells (y axis label: cell number) to a selection of images. Solid black line represents the mean response over three trials; shaded black areas represent the 95% confidence interval. Pink, brown, and blue traces represent the corresponding model responses according to the legend below and the models illustrated in (F). More example responses are shown in Figure S5.

(C) Analogous to (B) but for Mi1.

(D) L2 example of all cells' responses at a single point in time during an individual recording, with the image frame at the cells' receptive fields in the background. The response dynamics to the moving image is shown in the Video S1.

(E) Analogous to (D) but for Mi1.

(F) Illustration of the models. The model output of the simplest model is the image luminance over time at the receptive field centers of the cells (pink, Luminance model). The second model's responses are generated by filtering the image sequences with the spatiotemporal receptive field kernel (from the white noise analysis) for each cell (brown, RF model). The last model adds a static nonlinearity to the RF model (L-NL model, blue).

(G) Pooled responses of all L2 axon terminals (N = 5 flies, n = 389 cells) in all recordings to all images at all time points, plotted against the RF model prediction (gray scale hexagons). Data count denotes the number of data points in each bin on a log-axis. The green trace represents the average, and the green shaded area represents the 95% confidence interval (over binned data).

(H) Mean response as in (G) and a least-squares logistic function fit (dashed blue line). This function is the static nonlinearity applied to generate the L-NL model predictions.

(I) Cellular reliability (black), described as the Pearson's correlation coefficient r between different trials (for details see Transparent Methods), and correlation between the different models and the data (pink, brown, blue). Error bars denote the 95% confidence interval.

(J) Analogous to (G) but for Mi1 (N = 4 flies, n = 351 cells).

(K) Analogous to (H) but for Mi1.

(L) Analogous to (I) but for Mi1.

Video S1. Dynamic Responses of L2 and Mi1 to Moving Images, Related to Figure 5

The video shows example L2 and Mi1 responses to moving natural scenes. Normalized responses are plotted on top of the section of moving image that the fly is viewing. Original oblique plane recording (average over three trials) is shown on the bottom left, rotated such that the azimuth axes correspond approximately.

We next investigated the representation of natural scenes by Mi1, a cell that is one synapse downstream in the motion vision pathway, postsynaptic to lamina neuron L1, and provides direct input to the direction-selective cell type T4. Mi1 responds to brightness increments with an increase in calcium levels (Arenz et al., 2017, Behnia et al., 2014, Strother et al., 2017), which we confirmed by the white noise analysis (Figures S3 and S4). We then measured Mi1 responses to moving natural scenes as for L2 (Figures 5C and 5E and Video S1) and created predictions of three different models. The static nonlinearity for Mi1 shows a pronounced saturation for low and high luminances (Figures 5J and 5K). As for L2, all models are able to explain significant variance of the data, yet none of them is able to fully predict the explainable variance (Figure 5L). Thus, although Mi1 represents a dynamic natural stimulus to a large degree as determined by its static properties, additional, likely dynamic, features add to this representation that are not captured by either of the models derived from white noise analysis.

Overall, we could demonstrate that the retinotopic populations of L2 and Mi1 code for local luminance changes of static and dynamic natural scenes, with an additional dynamic component not described by linear or static nonlinear receptive field models.

Discussion

Here, we present a remote-focusing module for two-photon microscopy to image oblique planes with uncompromised resolution and frame rate. We used the technique to simultaneously image over 100 cells in the retinotopic plane of the Drosophila visual system, a significant portion of its visual field. This represents a large increase compared with the small number of simultaneously recorded cells (∼10) in similar preparations. Showing natural scenes to the cell types L2 and Mi1, we could demonstrate that the linear receptive field as measured with white noise can predict the cells' responses to natural scenes to a large extent.

In the original publication on remote focusing, and some in subsequently published adaptations, the remote focusing unit follows the lateral scan unit in the beam path, or is embedded between scan and tube lens (Botcherby et al., 2012, Colon and Lim, 2015, Rupprecht et al., 2016). Typical two-photon microscopes do not allow for additional optical elements between lateral scan unit and imaging objective. In this article, we presented a system in which the remote-focusing module is placed ahead of the lateral scan unit in the beam path. This way, standard two-photon microscopes can be equipped with this module without changes in the arrangement and design of the original optical elements. Sofroniew et al. (2016) also placed a remote-focusing module ahead of the lateral scan units in their system. Their lateral scan optics comprised three lateral scan mirrors and several relay telescopes, optimized for large field-of-view imaging, which makes it hard to compare the performance of their remote focusing with a remote-focusing module in front of a conventional two-photon microscope.

The distance between two columns in the visual system of the fruit fly is about 5 μm. Therefore, we aimed for a near-diffraction-limited focal spot size (0.5 × 0.5 × 3 μm) throughout the volume. A typical sine-corrected objective lens forms a diffraction-limited focal spot in the focal plane, but aberrations are introduced when focusing above or below the focal plane (Botcherby et al., 2008, Botcherby et al., 2007, Conrady, 1905). By introducing a second objective lens (the remote lens), aberrations are introduced on purpose to be then canceled out by the imaging lens. Full cancellation is achieved with a 4f telecentric system and an overall magnification of n2/n1 between the two objectives (Botcherby et al., 2007). In similar system (Rupprecht et al., 2016), in which either the telecentric relay or magnification criterion were omitted, the axial resolution depended on the axial position of the mirror, reaching 7 μm, and the relationship between mirror displacement and focus displacement was no longer linear. Depending on the size of the structures of interest, this resolution loss can be neglected. For small biological structures, however, a system with higher resolution is preferable. In our system, the relationship between mirror movement and focal spot displacement was linear, and the magnification was close to the ideal magnification (1.42 instead of 1.33), resulting in an overall better resolution. The remaining aberrations in the axial point spread function are likely due to the remaining inaccuracy in magnification, due to the short distance between tube lens and imaging objective lens, or may stem from alignment imprecision. For a magnification even closer to ideal, lenses in the remote-focusing module could potentially be chosen to precisely match the magnification of tube and scan lens.

Most other studies have used remote focusing to jump to different planes along the z axis. Our approach of tilting the imaging plane provides a simple and efficient solution to maximize information content in one, oblique plane (see also: Colon and Lim, 2015) without reducing the imaging rate. A comparable result could be attained by volume-imaging approaches using acousto-optic deflector systems (e.g., Grewe et al., 2010, Katona et al., 2012) or electro-tunable lenses (Grewe et al., 2011). Acousto-optic deflectors require beam stabilization and dispersion control (Iyer et al., 2003), and both would involve substantial modifications to the microscope setup. Thus, our system provides a simple, low-maintenance alternative that utilizes existing two-photon hardware. Although we used the module only to image oblique planes, arbitrary line scans or other 2D manifolds, for example, spherical surfaces, can be achieved without changing the hardware, by software extensions only.

We demonstrated the applicability of our technique by imaging the responses of the cell types L2 and Mi1 in the visual system of Drosophila to natural images, within the retinotopic plane spanned by their axon terminals. We first characterized their linear receptive fields and mapped their spatial distribution. As expected from the layout of the eye, receptive field locations formed a regular, hexagonal, and retinotopic grid, mapping the 2D visual space. We did not observe a systematic difference in receptive field structure across the range of locations we measured (70 × 90°, elevation × azimuth). The response properties of L2 to moving gratings, edges, flashes, and white noise stimulation have been characterized in detail (Arenz et al., 2017, Clark et al., 2011, Drews et al., 2020, Freifeld et al., 2013, Reiff et al., 2010), suggesting an OFF center, ON surround, and band-pass receptive field with an anisotropic structure, which we confirmed by white noise analysis. We also confirmed the receptive field of Mi1 as previously described (Arenz et al., 2017) and in addition identified a spatial anisotropy as for L2.

Our results show that the dynamic responses to natural scenes are largely, but not completely, predictable using classical, static linear-nonlinear models. This indicates dynamic nonlinear properties of the system. Lamina cells in blowflies have been shown to dynamically adapt to luminance (Laughlin and Hardie, 1978) and to natural, temporal sequences of luminances (van Hateren, 1997). The photoreceptors in Drosophila also adapt to and compress luminance levels (Juusola and Hardie, 2001), and similar phenomena have been described for other lamina neurons (Tuthill et al., 2014) and also recently for L2 (Ketkar et al., 2020). This dynamic adaptation could cause the variance not explained by the L-LN model for L2. Mi1 has recently been shown to exhibit spatial normalization to surround contrast, as well as temporal adaptation to variations in input contrast (Drews et al., 2020, Matulis et al., 2020). Contrast in natural scenes is highly variable, and these dynamic adaptation processes could therefore be responsible for the residual variance in Mi1.

Several studies in other model organisms have found both similarities and differences in visual response properties when comparing artificial with naturalistic stimuli. For example, complex cells of V1 in cats respond to natural scenes as predicted from their linear receptive field (Touryan et al., 2005), despite their known nonlinear properties (Hubel and Wiesel, 1962). Natural scenes have been used for reverse correlation analysis, which generally yielded receptive fields comparable to the ones obtained with white noise stimuli (Ringach et al., 2002, Theunissen et al., 2001, Touryan et al., 2005, Vance et al., 2016, Willmore et al., 2010). On the other hand, a recent study highlights the dynamic nonlinearities during natural scene stimulation in salamander retinal ganglion cells (Maheswaranathan et al., 2019), where white-noise-derived models are not able to explain their data well. Overall, whereas many response properties of cells can be probed with artificial stimuli, the use of natural stimuli can further our understanding of a dynamic neural circuit that has evolved in natural conditions.

In summary, with remote-focusing microscopy, we could effectively visualize the retinotopic maps in the Drosophila visual system. We showed that two early visual cell types form 2D feature maps of the fly's environment. We expect that our method will provide insights on spatial relationships of the cells in the visual system of the fly. Moreover, our module for imaging oblique planes can be implemented in any conventional two-photon system, which opens possibilities to investigate neural structures from new angles.

Limitations of the Study

The distance between the imaging objective and the tube lens f4 in conventional two-photon microscopes, including ours, is shorter than required for a complete 4f telecentric coupling between the remote objective and imaging objective lens. Therefore, our system could not be constructed entirely aberration-free. In addition, the magnification between remote objective and imaging objective lens is 1.42, whereas ideally it should be 1.33, which is another factor that introduces aberrations. Nevertheless, for the imaging volume required for our application, the aberrations are negligible. We are using gray scale natural images that are compressed in their luminance range, and that exhibit only global motion at constant speed. Although these images represent many parameters of the natural environment, they do not fully reproduce it. We are using the calcium sensor GCaMP6f to measure calcium signals of visual neurons. The dynamic properties of the sensor are far slower than the calcium dynamics in neurons, which limits the explanatory power of the measured signals, especially for the comparison between white noise and natural scene responses.

Resource Availability

Lead Contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact, Anna Schuetzenberger (schuetzenberger@neuro.mpg.de).

Materials Availability

This study did not generate new unique reagents.

Data and Code Availability

The modified version of ScanImage 5.1 created in this study is available at GitHub (https://github.com/borstlab/Scanimage_Planes3D). The datasets and the data analysis code generated during this study is available at GitHub (https://github.com/borstlab/obique-planes).

Methods

All methods can be found in the accompanying Transparent Methods supplemental file.