A versatile, multi-laser twin-microscope system for light-sheet imaging

Light-sheet microscopy offers faster imaging and reduced phototoxicity in comparison to conventional point-scanning microscopy, making it a preferred technique for imaging biological dynamics for durations of hours or days. Such extended imaging sessions pose a challenge, as it reduces the number of specimens that can be imaged in a given day. Here, we present a versatile light-sheet imaging instrument that combines two independently controlled microscope-twins, built so that they can share an ultrafast near-infrared laser and a bank of continuous-wave visible lasers, increasing the throughput and decreasing the cost. To permit a wide variety of specimens to be imaged, each microscope-twin provides flexible imaging parameters, including (i) operation in one-photon and/or two-photon excitation modes, (ii) delivery of one to three light-sheets via a trio of orthogonal excitation arms, (iii) sub-micron to micron imaging resolution, (iv) multicolor compatibility, and (v) upright (with provision for inverted) detection geometry. We offer a detailed description of the twin-microscope design to aid instrument builders who wish to construct and use similar systems. We demonstrate the instrument’s versatility for biological investigation by performing fast imaging of the beating heart in an intact zebrafish embryo, deep imaging of thick patient-derived tumor organoids, and gentle whole-brain imaging of neural activity in behaving larval zebrafish.


I. INTRODUCTION
Most of what we recognize as the phenomena of life are not properties of stationary structures but emerge from dynamic interactions among many elements over time. Modern optical microscopy methods offer an efficient means for non-invasive, high-resolution observation of many of life's most fascinating phenomena. 30 The difficulty is that light imaging involves unavoidable tradeoffs between spatial resolution, acquisition speed, field-of-view, penetration depth, and the limited photon budget from the sample. 1 Considerations of the photon budget are crucial to biological imaging, as there are a finite number of photons that a given fluorophore can emit before it is bleached, and high light doses on the specimen can lead to photo-induced toxicity. It is thus critical that the excitation light be used as efficiently as possible and the emitted photons be collected as efficiently as possible while striking the optimal imaging compromises for whatever the application demands.
Over the past decade, there have been a series of important developments in light-sheet microscopy, a century-old technique 2 (also known as selective-plane illumination microscopy; SPIM). 3 SPIM decouples the illumination and detection paths by using separate optics to excite and detect fluorescence: a cylindrical illumination lens is used to project a static, thin, two-dimensional (2D) sheet of light coincident to the focal plane of a detection objective lens. In contrast to confocal laser scanning microscopy and other point-scanning techniques that acquire volumetric information one voxel at a time, light-sheet excitation permits an entire 2D plane of fluorophores to be excited and detected with high signalto-noise ratio, high imaging speed, and minimal light exposure to the sample. Developments in light-sheet microscopy techniques [4][5][6][7][8][9][10][11] have led to cutting-edge applications across a range of fields [4][5][6][7]10,12,13 from developmental biology 5,14 to neuroscience. 10,12,[15][16][17] Each of these modifications of light-sheet imaging has involved their own tradeoffs in performance, complexity and expense of the FIG. 1. Light-sheet microscopy principle. A light-sheet (blue) can be created by dynamically scanning, along the y direction, a focused Gaussian beam that propagates in the x direction. The focusing is achieved via a low numerical aperture illumination lens. The fluorescence generated by the illuminated (x-y) plane is imaged (green) by an orthogonally positioned wide-field microscope. Axial (x-y) sections of the 3D sample are captured either by scanning the sample (orange) through the stationary focal plane or by scanning the light sheet and detection focal plane through the stationary sample.
imaging compared to 1P excitation. 5,9 Thus, for the best flexibility to handle a wide variety of samples, both 1P and 2P excitation modes are desirable. Because both use a low-NA illumination lens to generate an axially extended Gaussian focus, they can be conveniently carried out on the same light-sheet instrument. 19,21 A major expense of SPIM imaging setups is the laser source used for excitation, especially given that multiple lasers are used for multicolor imaging. For 2P-SPIM the ultrafast laser can more than double the equipment cost, which has limited its adoption, despite its superior performance in optically challenging samples. Even for well-funded laboratories, laser sources tied to a single microscope are not cost-effective. Since there is an upper limit to the amount of laser power that can be delivered to any specimen without perturbing it, most implementations waste well more than half of the total laser power available.
Another practical challenge comes from the often dual need of continuously imaging to generate time-lapses and of imaging a large number of samples to obtain statistically significant results. While the low photodamage of SPIM allows biological processes to be imaged for a duration of several hours or days, a tradeoff exists between such prolonged imaging sessions and the number of samples that can be imaged in a given day using a single instrument.
Here, we describe the flex-SPIM, which combines two independently controlled light-sheet microscope-twins that share an ultrafast NIR laser and a bank of continuous-wave (CW) visible lasers. This permits two specimens to be imaged simultaneously for far less than the cost of two multi-laser microscopes. Each microscopetwin has built-in modularity for tailoring its use on diverse samples and scientific questions. In Sec. II, we describe the flex-SPIM design in detail for those who wish to construct and use a similar instrument. In Sec. III, we test the performance against our design objectives by imaging three challenging specimens: the beating larval zebrafish heart, patient-derived tumor organoids, and whole-brain neural dynamics in behaving zebrafish. The results demonstrate the ease of adapting the flex-SPIM for application-specific light-sheet imaging.

II. INSTRUMENT DESIGN, INTEGRATION, AND CONFIGURATION
The flex-SPIM draws on lessons learned from proof-ofprinciple studies 19,21 with a new imaging technology to meaningful scientific results; [22][23][24] several years of interactions with end users at advanced imaging centers (at the California Institute of Technology and the University of Southern California), the 2P-SPIM inventors, and other instrument builders; and thus integrates the following combination of improvements: • Two independent microscope-twins share the same multilaser source (Fig. 2), dramatically reducing instrument cost-more than 30% savings. This also minimizes the ongoing maintenance and operation costs, since lasers need service and/or replacement after a finite lifetime. • The twin architecture doubles the 1P-and 2P-SPIM imaging throughput and increases the variety of specimens or imaging modes, when compared to a single microscope. • The system is "flexible" by design, with an opto-mechanical configuration that is both open and modular, providing 3D opto-mechanical model of the twin-microscope system mounted on a 5 × 10 ft 2 , anti-vibration optical table. Model shows the multi-laser subsystem shared between microscope-twin-1 (right) and microscope-twin-2 (left). Twin-1 has the four functional subsystems labeled and features an implementation of both upright and inverted detection. Brief descriptions of each functional subsystem are provided in Table I. The inset shows a detailed view of the sample chamber (SC), the dive bar (DB) used to hold the sample, three excitation objectives (EO) to deliver excitation light-sheets to the sample, and the detection objective (DO) to collect emitted fluorescence from the sample.
a straightforward path to instrument evolution and customization for different samples and applications. Three orthogonal illumination arms offer easy matching to different specimens, enhancing illumination uniformity or increasing optical coverage for larger and more opaque samples. 25 Switching from high lateral spatial resolution (submicrometer) to a lower resolution (∼micrometer) with a larger field-of-view requires only a simple adjustment to the detection subsystem. Each twin can further be configured in upright and/or inverted detection geometries to accommodate a diversity of specimens (Fig. 2).
The flex-SPIM consists of four functional subsystems (Fig. 2) and two modules (Table I) and sits on a 5 × 10 ft 2 , anti-vibration optical table (Fig. S1). The schematic diagram of the integrated illumination paths is shown in Fig. 3; the corresponding 3D

FIG. 3.
Schematic diagram of multi-laser and illumination-scanning optics subsystems of the instrument. Visible light from the continuous-wave laser bank is fed into microscope-twin 1 (right) and microscope-twin 2 (left) via polarization beamsplitting optics [consisting of a polarizing beamsplitter (PBS) and half-wave plate]. Acousto-optic tunable filters (AOTFs) are used to select the visible wavelengths and adjust the power independently for each twin. The near-infrared (NIR) light from the ultrafast laser is routed similarly using Pockels cells (PCs) to adjust the NIR power independently for each twin. The visible and NIR beams are raised onto 24 × 36 in. 2 optical breadboards by using periscopes (P). Polarization beamsplitting optics are used both to combine the visible and NIR beams and to split the combined beam into two paths (illumination arms 1 and 2). Illumination arm 1 is further split into two paths through polarization beamsplitting optics, creating a total of three illumination arms. Each illumination arm directs light to the sample through the excitation objectives (EO). BE-beam expander, M-mirror, DM-dichroic mirror, λ/2-half-wave plate, where the subscripts VIS and NIR refer to the visible and near-infrared wavelengths, respectively, G-2D scanning galvo mirrors, SL-scan lens, and TL-tube lens. BE in the NIR twin-2 path appears gray because it is underneath the optical breadboard. opto-mechanical model is shown in Fig. 2 and the complete computer-aided design (CAD) is available upon request. Whenever possible, commercially available hardware components are used; however, both basic machining of off-the-shelf parts and the fabrication of custom components are required (Table SI). Most standard optical elements are mounted in Thorlabs 30 mm or 60 mm cage components. Beam steering mirrors shared by both the ultrafast and CW lasers (illumination-scanning optics) have protected silver coatings, whereas those used by the ultrafast or CW lasers alone have broadband dielectric coatings.
A. Multi-laser subsystem CW visible light used for single-and multi-color imaging via linear excitation is provided by a bank of CW lasers (445 nm, 488 nm, 561 nm, and 647 nm), collimated and expanded to a 1/e 2 diameter of 1.5 mm, and combined into a co-linear beam using broadband and dichroic mirrors (see Table SI). 26 The combined beam is then split into two paths of equal length and power through polarization beamsplitting optics [consisting of a half-wave plate mounted on a rotation mount and polarizing beamsplitter (PBS)], delivering light to each microscope-twin. Each light path passes through acousto-optic tunable filters (AOTFs), which are used to select the wavelengths and adjust the power independently for each twin. The AOTFs (Table SI) require the input laser beam polarization to be linearly orthogonal to the baseplate (s-polarization) to maximize the diffraction efficiency and ensure chromatic colinearity of the modulated beam. Alternatively, the input beam can be p-polarized if the AOTF crystal output face is used as the "input" face due to the Helmholtz reciprocity principle. 27 Because of the upstream polarizing beamsplitter used, the beams are p-polarized for the twin-1 path and s-polarized for the twin-2 path. As such, the AOTF in the twin-1 path is mounted so that the p-polarized beam enters the AOTF through its output face; the AOTF in the twin-2 path is mounted conventionally. A half-wave plate can be placed in front of the AOTF to fine-tune the polarization direction of the beam entering the AOTF and thereby maximize the diffraction efficiency by the AOTF when more excitation energy is required at the sample, as we have implemented in the twin-2 path (Fig. 3). Note that alternatively, similar performance could be achieved by placing a half-wave plate in front of the AOTF in the twin-1 path to rotate the beam's polarization so that it enters the AOTF conventionally.
The tunable NIR ultrafast mode-locked laser used for singleand multi-color imaging via nonlinear excitation is split into two paths of equal length and power through polarization beamsplitting optics, delivering NIR light to both microscope-twins. Note that the linear polarizations of the beams for the twin-1 path and the twin-2

ARTICLE scitation.org/journal/rsi
path are orthogonal to each other. Each path passes through a Pockels cell to control the power independently for each twin (Fig. 3). Each Pockels cell is rotated to match its input polarization requirement and hence maximize its extinction ratio, removing the need for additional half-wave plates and power losses from their imperfections. Following the Pockels for each path, the beams are expanded to a 1/e 2 diameter of 2.2 mm. A long-pass filter (800-nm cutoff) is mounted upstream of the beam expander to block any undesired residual wavelengths from the ultrafast laser source.

B. Illumination-scanning optics subsystem
Visible and NIR beams from the multi-laser subsystem are each raised onto a pair of 24 × 36 in. 2 optical breadboards, one for each twin, by periscopes (Fig. 3). Polarizing beamsplitters are used to both combine the visible and NIR beams into a co-linear beam and to split the combined beam into two paths (illumination arms 1 and 2). Visible and NIR half-wave plates, each mounted in manual rotation mounts, are used to adjust the laser power (splitting ratio) delivered to illumination arms 1 and 2. Illumination arm 1 is further split into two paths through polarization beamsplitting optics, creating a total of three illumination paths. Rotation of the NIR half-wave plates before the polarizing beamsplitter can be used to adjust the relative laser power into illumination arms 1 and 3 (Fig. 3). The path lengths of all three illumination arms are routed so that they are equal.
The beams from each illumination arm are sent to a 2D scanning galvo mirror positioning system. The first galvo mirror rapidly scans the beam laterally to synthesize the light-sheet (in the x-y plane), and the second galvo mirror, which is conjugate to the back pupil of the excitation objective lens, translates the virtual light-sheet along the (z) detection axis. Following the scanning system, each illumination beam passes through a scan lens (achromatic doublet; see the supplementary material Note 1); a tube lens, and a low magnification, low-NA, long-working-distance excitation objective lens. The distances between pairs of lenses form a 4f arrangement (the distance between pairs of lenses are equal to the sum of their focal lengths).
Three excitation objective lenses, mounted orthogonally to each other, direct the illumination light toward the sample (Fig. 3). Depending on the sample properties, any combination of the excitation objectives can be used, either sequentially or simultaneously. Small and/or transparent samples, for example, may benefit from single-sided illumination with 1P excitation, whereas relatively large and thick samples may benefit from the uniform illumination coverage offered by using all three objectives with 2P excitation.
The illumination NA for 1P and 2P is ∼0.02 and 0.03, respectively, yielding fluorescence Gaussian-beam light-sheets of ∼10 μm in average thickness across an extent of ∼400 μm, as shown in Fig. 7(d). Note that to produce similar nominal fluorescence lightsheet thicknesses and extents, different NAs are needed for 1P vs 2P due to the difference in laser wavelength and the linear (1P) vs quadratic (2P) dependence of the fluorescence on the laser intensity. 19 Scanning of the first galvo yields an effective (x-y) field-ofview of ∼400 × 1000 μm 2 . The chosen light sheet thickness ensures that we are able to resolve single neurons (6-8 μm in size) throughout the entire ∼400 × 800 × 250 (x-y-z) μm 3 brain of zebrafish larva at 5 days-post-fertilization (dpf). 20,28 Light throughput for each twin, defined as the total measured laser power at the sample from the three illumination arms divided by 50% of the measured power at the laser output (for 1:1 split into each twin) for the ultrafast laser (taken at λ = 900 nm) is ∼60%. Light throughput in the visible regime, taken by throughput measurements across each CW laser line is ∼6%. The lower visible light throughput is expected since most of the illumination-scanning optics were selected to optimize NIR throughput and maximize 2P excitation efficiency (Table SI). Both the ultrafast and CW laser sources are able to simultaneously run experiments on both twins with independent power control.

C. Detection subsystem
The sheet of fluorescence signal generated at the sample is collected by an orthogonally positioned water-immersion detection objective lens (20×, 1.0 NA; see Table SI), mounted to a piezoelectric (piezo) collar. The high-NA objective, with low intrinsic magnification, not only enables high-resolution imaging across large volumes but also importantly maximizes light-collection, which is critical for maintaining acceptable signal-to-noise while minimizing the excitation laser power to reduce photodamage to live samples. The fluorescence signal passes through a filter wheel equipped with emission filters to block the excitation light and transmit the fluorescence signal emitted by the sample; the emission filters are optimized for the transmission of common fluorophores spanning the visible spectrum (Table SI). A tube lens forms the primary image of the fluorescence signal onto a scientific complementary metal-oxide-semiconductor (sCMOS) camera.
Different magnifications are achieved with the same detection objective by using tube lenses of different focal lengths, with the rail-mounted fluorescence camera providing the necessary distance adjustments (Figs. 2 and 4). Generally, tube lenses need to have a large enough diameter to capture the detection optical aperture and good performance to minimize chromatic and geometric aberrations. 30 Lower or higher magnification requires lenses with shorter or longer focal lengths, potentially exacerbating aberrations or making the instrument footprint larger, respectively. We find that the practical range of magnification is between 44× and 11× for our instrument (Table SI), given the available off-the-shelf double achromats (used as tube lenses), the size constraint of standard opto-mechanical components, and the desire to keep our instrument reasonably compact.
Our detection subsystem design choices described above represent a balanced compromise to achieve an economical light-sheet instrument with adjustable magnification, maximizing resolution while maintaining a large field-of-view. As will be shown later, the instrument provides cellular to sub-cellular resolution, sufficient for a wide range of applications. We note that the maximum diffractionlimited resolution, as determined theoretically by the NA of the detection objective, is not expected to be achieved, given the practical design choices that we have made. First, at the magnifications used, the camera pixel size is not small enough to provide diffractionlimited spatial sampling according to the Nyquist criterion. 30 Second, independent of the magnification used, we do not expect our detection path to be fully aberration-free due to our choice of using economical off-the-shelf double-achromats as tube lenses, rather than more expensive specialized tube lenses (which often only exist with limited focal lengths). If higher resolution is desired with our system, a detection objective with higher NA and/or higher intrinsic magnification could be used in combination with the matched tube lens from the same manufacturer. Depending on the sample properties, the detection subsystem can be arranged for upright and/or inverted configurations. We have experimentally implemented an upright configuration (Fig. 4), as it is the most optimized for the biological samples presented here. We implemented an inverted configuration in the CAD model shown in Fig. 2, where the same detection subsystem was designed to be mounted in either an upright or inverted configuration. Owing to the system's arrangement of opto-mechanical components (Figs. [2][3][4], switching between detection geometries or changing the overall magnification is relatively straightforward.

D. Sample mounting and motion control subsystem
The sample chamber has three side windows for the excitation objectives as well as a bottom window to provide an additional view of the specimen. The sample chamber is open at the top and is filled with imaging buffer; the open-top allows the detection objective to be liquid-immersed and the sample holder to be inserted. The sample chamber sits on a custom heat exchanger that has circulatory channels for temperature regulated fluid flow, which can be used to keep the medium-filled sample chamber at a specific temperature.
The sample holder is comprised of two parts: (i) a caddy that holds the specimen and (ii) a "dive bar" that holds the caddy and connects it to the stack-up of motion-control stages. Caddies can be used for agarose embedding of the sample or adapted to specific applications. For example, the caddy for imaging neural activity in behaving zebrafish [ Fig. 5(b)] immobilizes the specimen's head with 1.5% low-melting agarose gel (to record cellular resolution wholebrain neural activity), leaving the tail free to move (to record swimming behavior). With one end of the dive bar holding the caddy immersed in the imaging buffer, the other end is mounted to a dual-axis goniometer, providing rotational motions around the xand y-axes [ Fig. 5(b)]. The goniometer is in turn mounted on a 3D stack-up of motorized linear stages (see Table SI), with each stage providing ±25 mm of travel range. The combination of the two-axis goniometer and 3D stage stack-up allows fine sample positioning so that the illuminated region of interest can be overlapped with the detection objective focal plane. The two-part design of FIG. 5. SPIM imaging of intracellular calcium for capturing neuronal activity. (a) Schematic of apparatus for imaging of neural activity during various behaviors in the larval zebrafish. Sheets of laser light are synthesized by quickly scanning the pulsed illumination beam (red) with galvo mirrors (G). 2P light-sheets are delivered to the agarose-embedded head of the animal with excitation objectives (EO) from the side and front arms. The side masks cover each eye on the sides of a horizontally oriented zebrafish, while the front mask covers both eyes, enabling access to neurons between the eyes. 2P-excited calcium fluorescence signal is collected through an upright detection objective (DO) and onto a scientific CMOS camera. A triggerable wide-field camera is positioned below the sample chamber (SC) to provide a wide-field, low-resolution view of the sample, as shown in (b). During a typical neural imaging experiment, the zebrafish larva is mounted in a caddy, which in turn is mounted to the dive bar (DB) underneath the DO. Within the caddy, the zebrafish's head is immobilized in agarose, while the tail is free, permitting the monitoring of zebrafish behavior through tail movement. SL-scan lens, TL-tube lens, SC-sample chamber, and L-camera lens. The third illumination arm, emission filter, detection TL, scientific camera, light-emitting diode, and filter for behavior channel are not shown. Insets in (b) highlight that the calcium fluorescence channel (green) is recorded from the zebrafish brain, while the behavioral channel (dark red) monitors the tail movement of the animal.  the sample holder allows for the flexibility of designing different caddies to be best suited for specific types of specimens while still using the same dive bar to connect the caddy to the motion-control stages.

Review of
The flex-SPIM has two different modes of capturing volumetric information from a 3D sample: either by sample-scanning or by objective-scanning.
• Sample-scanning mode: the excitation light-sheet and detection objective remain stationary; the sample is moved via the z-stage of the 3D stage stack-up along the optical axis of the detection subsystem and images are sequentially collected. As the z-stage is already a necessary part of sample positioning, this approach is the simplest and most economical to implement. The motion range of z-stages is typically tens of mm; hence, sample-scanning enables volumetric imaging of large samples with depths up to the working distance of the detection objective (several mm or more). However, the imaging speed is limited by mechanical inertia of the stage and sample holder and by the communication overhead between the acquisition computer and the z-stage controller. Furthermore, the translational motion of the specimen can compromise normal biology.

ARTICLE scitation.org/journal/rsi
• Objective-scanning mode: the movement of the detection objective piezo collar is synchronized with the second galvo mirror of each illumination arm-the light-sheet and piezo collar are scanned axially in concert (with a travel range of ±500 μm set by the piezo collar) and images are sequentially collected (further details are described in Sec. II E). This approach, notwithstanding the additional cost of the piezo collar and its limited range of motion, enables fast volumetric imaging without moving the specimen and is preferred for our whole-brain functional

Review of Scientific Instruments
ARTICLE scitation.org/journal/rsi imaging and simultaneous behavioral observation studies with zebrafish.

E. Instrument control module
Each twin is independently controlled with a computer equipped with two Xeon E5-2650 v4 processors and 128 GB of 2400 MHz DDR4 RAM and seven PCIe slots, enough space for all the control cards. Instrument control and image acquisition are done through Micro-Manager. 29 Custom software developed in Lab-VIEW (National Instruments) is used to independently control each of the 2D scanning galvo mirror systems and allows precise alignment (size and swept-rate control) of the excitation light-sheet relative to the sample (software available upon request). Collected images are written directly to a dual-disk array consisting of eight 7200 rpm, 4-TB disks.
In the objective-scanning mode, the piezo collar's controller serves as the master timing source. The analog position-readout of the piezo collar triggers a PicoScope, which is used to generate control signal sequences to synchronize the camera(s) with image capture. The position-readout of the piezo collar is also used to drive the position of the z-galvos. The waveform from the timing output of the scientific CMOS camera controls the AOTF and Pockels so that the sample is not illuminated during the camera readout. Analog control signals for the galvos and Pockels are appropriately conditioned by individual scaling amplifiers. A schematic of the control signal sequences is shown in (Fig. 6).

F. Auxiliary module
Illumination can be selectively blocked with masks to avoid photosensitive regions or autofluorescent features in samples. For example, to avoid illuminating the zebrafish eyes while imaging neural activity, the excitation light is physically blocked with a pair of masks on each side of a horizontally positioned zebrafish and another mask for the front that covers both eyes [ Fig. 5(a)]. These masks, fabricated out of black anodized aluminum, are mounted at the image planes of the illumination-scanning optics, each on 2D translational stages to permit their accurate positioning for different specimens or their complete removal from the illuminated field. A far-red light-emitting diode and a wide-field camera, positioned to view the sample from the bottom, enable view-finding and monitoring the tail behavior during neural imaging of the zebrafish (Fig. 5).

III. INSTRUMENT PERFORMANCE
We characterized the 3D resolution of the flex-SPIM by measuring the point-spread function (PSF) with sub-diffraction fluorescent beads and then demonstrated the utility of the instrument for investigating biological systems by imaging the beating embryonic zebrafish heart, thick patient-derived tumor organoids, and neural activity in behaving zebrafish.

A. Resolution characterization
We measured the system PSF with sub-diffraction (175 ± 5 nm diameter) fluorescent beads (PS-Speck Microscope Point Source Kit, P7220, Molecular Probes) embedded in 1.5% agarose, using the same detection objective (20×, NA = 1.0) with different tube lenses to yield two different magnifications of 44× and 11× (Table SI). Representative bead images and quantitative values of the PSF are presented in Fig. 7. On the whole, the instrument achieved ∼0.5 μm and 1 μm lateral resolution, for 44× and 11×, respectively, and ∼1.8 μm axial resolution at both magnifications. This performance comes close to but does not achieve the full theoretical resolution limit determined by the NA of the detection objective, which is expected due to the practical compromises that we have made with the detection subsystem, as discussed previously in Sec. II C. Overall, the demonstrated resolution is quite suitable for a variety of imaging applications and represents the practical compromise between achieving a high resolution while maintaining a large field-of-view. The vertebrate heart is a highly dynamic organ that starts to take its form and function early on during development. 32 To gain insight into how the heart develops, studies of cells in their native dynamic and 3D context in the intact heart are needed. While the zebrafish is an ideal model system because of its optical and genetic accessibility, 33 imaging is challenged by the heartbeat motion at 2-4 Hz and over tens of micrometers in amplitude. 34 Retroactive synchronization techniques can align the 2D images by taking   FIG. 12. Whole-brain functional imaging at single-cell resolution in behaving 5-dpf transgenic larval zebrafish expressing nuclear-localized calcium indicator elavl3:H2B-GCaMP6s. (a) Maximum-intensity projections of calcium activity are color-coded in time over the 30-min recording window. Active neurons that exhibit fluorescence change during the recording appear as colored dots. Volume of 400 × 800 × 250 (x-y-z) μm 3 was sampled by 52 z-planes (4.8 μm z-steps) at 0.5 Hz and 11× magnification. Figure 13 (Multimedia view) shows a 3D rendered movie of the same specimen. (b) Magnified volume renderings of neuron 1 (left) and neuron 2 (right) outlined in the x-y projection in (a), demonstrating cellular resolution in whole-brain 2P-SPIM imaging. The standard deviation projection along the temporal axis was used to generate the renderings, and a false-color (hot) lookup table was used to aid visualization. (c) Representative single-neuron activity traces extracted from the whole-brain recordings, calculated as ∆F/F from the manually selected neurons in (a). (d) Plot shows the total sum intensity of the entire imaged volume as a function of time, normalized to the average intensity of the first 5 min of acquisition (gray line at y = 1). Total intensity exhibits less than a 0.5% decrease in magnitude after the 30-min recording window, indicating that photobleaching is negligible. (e) Analysis of tail flick behavior as a function of time. Tail flicks (movements) obtained by taking the absolute difference in the pixel intensities from the tail region of the wide-field view of the sample, as shown in Fig. 5(b), normalized by the average image intensity corresponding to when there is no tail movement (gray lines). A threshold was applied to classify tail flicks and subsequently confirmed by manual inspection. Tail  Review of Scientific Instruments ARTICLE scitation.org/journal/rsi advantage of the quasi-periodicity of the heart motion. 21,31,35,36 We acquired 2D images of the beating heart at a high-spatiotemporal resolution in a 5-dpf transgenic larval zebrafish with the vasculature fluorescently labeled. The 1P excitation (λ = 488 nm) light-sheet was parked in z to optically section through the beating heart of the agarose-embedded sample as we acquired images at 85 frames/s, 11× magnification, and subcellular resolution [Figs. 8 and 9 (Multimedia view)], showing that the flex-SPIM is fully compatible with existing retroactive synchronization techniques. 21,31,35,36 C. Imaging of thick, patient-derived tumor organoids 3D cell culture systems, such as spheroids or organoids, recapitulate the native physiology of multicellular tissues much better than 2D culture systems. 37 Multicellular cancer organoids permit the study of disease development and patient-specific response to therapy. 38,39 Unfortunately, such multicellular systems are scattering and aberrating, making them challenging to image with conventional instruments.
To show the advantages of the flex-SPIM for such opaque and optically heterogeneous samples, we imaged, at 11× magnification, chemically fixed, agarose-embedded organoids differentiated from cells derived from a colorectal cancer patient that had been engineered to transgenically express nuclear-localized H2B-GFP [Figs. 10(a) and 11 (Multimedia view)]. 2P-SPIM provides better contrast throughout the imaged volume because the reduced scattering at the longer wavelength (λ = 900 nm) enables betterpreserved light-sheet shape over longer propagation distances compared to 1P [Figs. 10(b), 10(c), and 10(e)]. Even when the excitation light scatters, the fluorescence signal is still spatially restricted mainly to the central part of the light-sheet (where intensity is the highest) because of the quadratic dependence of the 2P-excited fluorescence signal on the excitation intensity. Thus, by mitigating the scattering-induced thickening of the light-sheet, 2P excitation with the flex-SPIM yields better effective 3D resolution than 1P (Fig. S2) and captures better images of labeled cells deep in the specimen [Figs. 10(d) and 10(f)].
D. Whole-brain functional imaging of behaving zebrafish SPIM enables recording of whole-brain neural activity in transgenic larval zebrafish. 15,16 These implementations, however, potentially stimulate the photoreceptors and other photosensitive cells in the retina with the visible excitation wavelengths used during acquisition. Such illumination can reduce visual sensitivity to stimuli and interfere with visually driven processes. 20 NIR (λ = 930 nm) 2P-SPIM overcomes this problem, 20 achieving a recording depth of 120 μm at a 1 Hz volume rate (sampled by nine z-planes). 17 We push the depth of 2P light-sheet functional imaging further with the flex-SPIM: more than doubling the volume size while maintaining high spatiotemporal performance [Figs. 12 and 13 (Multimedia view)]. By employing a trio of 2P excitation arms with masks to avoid direct laser illumination to the animal's eyes, we imaged the entire (400 × 800 × 250 μm 3 ) brain of a 5-dpf zebrafish expressing a pan-neural calcium indicator (elavl3:H2B-GCaMP6s) 16 at a 0.5 Hz volume rate (sampled by 52 z-planes) with single-neuron resolution and simultaneously monitored swimming behavior with a wide-field FIG. 13. Dorsoventral (left) and rotating (right) maximum-intensity projections of a time-lapse recording of the whole-brain of the a 5-dpf transgenic larval zebrafish. Two-photon whole-brain functional light-sheet imaging was performed at a volumetric rate of 0.5 Hz. The video loops a 5-min recording as part of the data presented in Fig. 12 (Fig. 5). The temporal and spatial resolution of the timeseries were sufficient to visualize individual active neurons across the brain [Figs. 12(a) and 12(b)] and capture single-neuron activity traces by calculating the relative fluorescence variation (∆F/F) as a function of time [ Fig. 12(c)]. Imaging was carried out for 30 min continuously with 490 mW of total average laser power delivered to the sample from the three illumination arms. We observed no apparent signs of phototoxicity from the animal's macroscopic behavior [ Fig. 12(e)], cellular structure [Figs. 12(a) and 12(b)], or calcium dynamics [ Fig. 12(c)]. Furthermore, the total fluorescence signal decreased by less than 0.5% after the 30-min, 900-z-stack acquisition [ Fig. 12(d)], indicating that photobleaching was negligible. In future work, systematic analysis of the 4D functional imaging data acquired by the flex-SPIM, aided by state-of-the-art analysis pipelines such as CaImAn, 40 will enable quantification of whole-brain singlecell activity patterns to gain insights into the neural basis of behavior.

IV. DISCUSSION
We present the design and construction of the flex-SPIM, an instrument with two independently controlled light-sheet microscope-twins sharing the same multi-laser source, dramatically cutting the cost of the system. We demonstrate instrument versatility and application-specific customization by imaging a variety of specimens. In the same spirit as the OpenSPIM project, 41 we offer

ARTICLE
scitation.org/journal/rsi a blueprint for optical developers to build and/or modify the flex-SPIM to serve user needs. Our design choice of upright (or inverted) detection, combined with our sample-mounting scheme using the caddy and the dive bar, allows for the flexibility in mounting different types of samples, from fixed tissues to live zebrafish embryos, but makes our instrument not easily amenable to multiview imaging, 42 where the sample is imaged from multiple directions to improve axial resolution and coverage. On the other hand, a number of other modifications and enhancements could be implemented on the flex-SPIM to further optimize its performance for specific applications. Incoherent structured-illumination from intensity-modulated illumination patterns generated by the AOTF and/or Pockels cell would enhance contrast in more scattering specimens but would require additional exposures and post-processing. 43 Confocal line detection using the rolling shutter of the sCMOS camera is an efficient alternative to structured-illumination and would allow rejection of non-ballistic photons. 44 Designing a sample chamber rig with both temperature and CO 2 control in conjunction with inverted detection would allow for optimized live imaging of cultured tissues as well as organoids.
Light-field microscopy could be readily deployed on the flex-SPIM, enabling high-contrast, synchronous volumetric imaging with SPIM-inspired selective-volume illumination. 45 In addition, implementing multispectral imaging would improve signal multiplexing, either on the illumination path by rapid multispectral excitation 46 or on the detection path by de-scanned detection via a confocal slit and diffraction grating. 47 Further improvement is possible with our hyperspectral phasor software (HySP) for unmixing multiple spectrally overlapping fluorophores, even in the face of low signal-to-noise. 48 The combination of HySP with a multispectral flex-SPIM design could thus enable dynamic visualization and quantitative analysis of many more important components and their interactions in intact specimens at high-resolution over extended durations.

V. ETHICAL APPROVAL STATEMENT
All zebrafish raising and handling procedures followed guidelines established in the Guide for the Care and Use of Laboratory Animals by the University of Southern California, where the protocol was approved by the Institutional Animal Care and Use Committee (IACUC). All zebrafish lines used are available from ZIRC (zebrafish.org).

SUPPLEMENTARY MATERIAL
See the supplementary material for a detailed list of the main flex-SPIM parts (Table SI), a panoramic photograph of the flex-SPIM (Fig. S1), optical transfer function comparisons of patientderived tumor organoids (Fig. S2), and simulations of scan lens performance (Supplementary Note 1).