The ideal fluorescence microscope for large-scale tissue imaging would provide: (1) nanoscale resolution, (2) over centimeter-scale volumes, (3) with minimal tiling and sectioning, (4) high isotropy in resolution and contrast, and (5) fast imaging speed. Multiple imaging systems have been developed within this space, each with unique advantages but inevitable trade-offs (Huisken et al., 2004; Dodt et al., 2007; Wu et al., 2013; Kumar et al., 2014; Tomer et al., 2014; Economo et al., 2016; Narasimhan et al., 2017; Power and Huisken, 2017; Migliori et al., 2018; Chakraborty et al., 2019; Voigt et al., 2019; Voleti et al., 2019; Chen et al., 2020b; Chen et al., 2020a; Glaser et al., 2022; Wang et al., 2021; Xu et al., 2021; Zhang et al., 2021; Qi et al., 2023; Vladimirov et al., 2024).

The choice of microscope objective is a critical design choice. Microscope objectives impose trade-offs between the smallest objects that can be resolved (i.e. the resolution), how much of the specimen can be observed at once (i.e. the field of view), and how thick of a specimen may be imaged (i.e. the working distance). These trade-offs are not based on physical law but reflect limitations of optical design, engineering, and lens manufacturing.

The trade-off between resolution and field of view is related to the etendue (G), which is proportional to the number of resolution elements of an optical system. The etendue is a quadratic function of the lens field of view (FOV) and numerical aperture (NA).(1)G=π4(FOV×NA)2\begin{document}$$\displaystyle G = \frac{\pi}{4}(FOV\times NA)^2$$\end{document}

Similarly, for a given NA, the required lens diameter increases proportionally with the required working distance (WD). The ideal lens for large-scale volumetric microscopy would provide large G and WD values.

90% of commercially available objectives for biological microscopy have G<1 mm² (Zhang and Gross, 2019a), and the vast majority of commercially available lenses have lens diameters < 2 cm. Several attempts have been made to develop custom lenses for specific applications. The ‘Mesolens’ provides an etendue of G=6.25 mm², with FOV = 6 mm, NA = 0.47, and WD = 3 mm (McConnell et al., 2016), but difficulties with manufacturing have limited adoption (McConnell, 2020). An objective with an etendue of G=7.07 mm², FOV = 5 mm, NA = 0.6, and WD = 2.7 mm has been developed for two-photon microscopy in vivo (Sofroniew et al., 2016). However, this lens is customized for infrared illumination and exhibits significant field curvature. Additional custom lenses have subsequently been developed for two-photon microscopy, although none of these lenses are suitable for large-scale fluorescence microscopy (Ota et al., 2020; Rumyantsev et al., 2020; Yu et al., 2021). A custom lens with a large etendue of G=18.86 mm², FOV = 14 mm, NA = 0.35, and WD = 19 mm has been developed for the real-time ultra-large-scale high-resolution (RUSH) microscope (Fan et al., 2019). Despite the large etendue, the lens is not commercially available and not designed for immersion, which is critical for imaging cleared or expanded tissues. Custom lenses such as the Schmidt (Voigt et al., 2024), Cousa (Yu et al., 2024), and Kyocera (https://www.ksoc.co.jp/en/seihin/immersion-objective/immersion-objective.html) lenses offer long working distances but small etendues. See Appendix 1—table 1 for a summary of the existing custom microscopy objectives.

Rather than designing a highly customized lens from scratch, we leveraged engineering investments that have been made in a different domain. High-resolution, high-speed imaging is widely used within the machine vision and metrology industry, where optical microscopes are used to map defects in semiconductors and other electronic devices. As the physical size of electronic components (e.g. pixels on flat panel displays) has become smaller, lenses for this domain have been designed with increasing NA, accessing biologically relevant resolutions (<1 µm). For a given NA, these lenses have remarkably large fields of view (and thus high etendues), low-field curvature, minimal distortion, and chromatic correction throughout the visible wavelengths. A comparison of the etendue as a function of NA and working distance for these lenses, and commercial and custom life sciences lenses is summarized in Figure 1. A more detailed summary of many electronics metrology technologies is provided as Supplementary file 1. Despite their superb specifications, both the lenses and camera sensors are readily available, are manufactured in large quantities, and are cost effective. We investigated the performance of metrology optics and cameras for imaging biological tissues and adapted a particular set of machine vision optics for biological microscopy.

The detection path of the system uses a lens with 5.0×magnification, with diffraction-limited imaging at NA = 0.305, and a 16.8 mm field of view (G=19.65 mm²) (VEO_JM DIAMOND 5.0×/F1.3, Schneider-Kreuznach, Germany) (Figure 2a). A 35-mm-thick glass beam splitter, normally used for co-axial illumination, is replaced with an optically equivalent thickness of liquid media. This enables spherical aberration-free imaging through 35–40 mm of liquid media (tunable for refractive indices between 1.33 and 1.56), including expanded hydrogels and cleared tissues. The lens is paired with a large-format CMOS sensor (VP-151MX, Vieworks Korea), based on the Sony IMX411 sensor, with a single-sided rolling shutter, 151 megapixels, 14192 (H)×10,640 (V), and a pitch of 3.76 µm. The camera captures a field of view of 10.6×8.0 mm (13.3 mm diagonal), capable of covering an entire 3×expanded mouse brain in only 15 tiles (Figure 2b). A conventional SPIM system would require 400+tiles to image an entire cleared brain at equivalent resolutions (Figure 2—figure supplement 1). A comparison of the selected lens and camera sensor with a state-of-the-art cleared tissue objective lens (Nikon 20×GLYC) and sCMOS camera (Hamamatsu Orca BT-Fusion) is shown in (Figure 2—figure supplement 2).

To improve axial resolution, our system synchronizes an axially-swept light sheet with the rolling shutter of the Sony IMX411 sensor (Dean et al., 2015). The CMOS sensor parallelizes readout across 14192 pixels per row, thereby achieving equivalent or greater pixel rates than typical scientific CMOS (sCMOS) sensors, even with an increased time per line and a relatively low frame rate (<6.4 Hz). As a result, the ExA-SPIM microscope operates at a higher imaging speed with comparable signal-to-noise ratio (SNR) to sCMOS-based systems (Figure 2e). See (Figure 2—figure supplement 3) for noise comparisons between the Sony IMX411 and sCMOS sensors. The low frame rate of the sensor facilitates accurate axially swept imaging using generic scanning hardware at an imaging speed of 946 megavoxels/sec. The excitation lens (1–290419, Navitar) is infinity-corrected (to enable axially swept excitation) and provides diffraction-limited resolution at NA = 0.133 over a 16 mm field of view. When fitted with a custom dipping cap, the excitation lens provides a working distance of 53 mm in water. The beam shaping in the excitation path is configured to deliver a light sheet with NA = 0.10 and width of 12.5 mm (full-width half-maximum). A full CAD layout of the ExA-SPIM system is shown in (Figure 2—figure supplement 4). Due to the large light-sheet, the ExA-SPIM system uses high excitation laser powers (1000+mW; Figure 2—figure supplement 5), resulting in light intensities typically used in SPIM microscopy. Photobleaching is not detrimental at these imaging conditions (Figure 2—figure supplement 6).

Based on these lenses, and the travel limits of a mechanical motorized stage, the ExA-SPIM is capable of imaging a 200×52 × 35 mm³ volume, with ~1.5 µm lateral and ~3 µm axial native optical resolution, and minimal field curvature quantified using the methodology described in Vladimirov et al., 2024 and distortion (Figure 2c–d). We leveraged this large volume to image intact, expanded tissues.

With 3×tissue expansion, the system can image a native tissue volume of 67×17 × 12 mm³ with an effective optical resolution of ~0.5 µm laterally and ~1 µm axially. This amounts to >100 teravoxels, which can be captured without the need for physical sectioning and with minimal tiling. The large field of view of the ExA-SPIM system can also permit imaging cleared tissues, such as mouse brains, in a single tile (Figure 2—figure supplement 7). A summary plot comparing the focal volume, isotropy, and imaging speed of our new ExA-SPIM system to other large-scale volumetric imaging systems is shown in (Figure 2—figure supplement 8). Individual data points are listed in Appendix 1—table 2.

The large-scale and high-speed imaging made possible by the ExA-SPIM system required new software for controlling the microscope, as well as downstream computational pipelines for compressing and handling the resulting datasets (Figure 2—figure supplement 9). We developed custom acquisition software capable of interfacing with the VP-151MX camera driver (https://github.com/acquire-project), which streams imaging data directly to next-generation file formats (Beati et al., 2020; Clack et al., 2023; Moore et al., 2021). Using a combination of high-speed networking, fast on-premises storage, and real-time compression, our current imaging pipeline enables a throughput to cloud storage of >100 TB per day, using commodity hardware. This pipeline is actively being developed, with ongoing efforts focused on standardizing every processing step around the OME-Zarr file format (Moore et al., 2023).

Tissue expansion for microscopy (ExM) enables imaging optically transparent specimens at effective resolutions well below the diffraction limit of light microscopes (Chang et al., 2017; Chen et al., 2015; Tillberg and Chen, 2019; Figure 2—figure supplement 10). Moreover, ExM can produce optically clear specimens with low fluorescence background. Multiple ExM variations have emerged, driven by specific biological questions. These include engineered hydrogel chemistry for post-expansion molecular interrogation of proteins or RNA (Asano et al., 2018; Cho and Chang, 2022; Tillberg et al., 2016; Cui et al., 2022), formulations that provide gel stiffness (Chen et al., 2021) or tunable expansion up to >10 × (Damstra et al., 2022; Klimas et al., 2023; Sarkar et al., 2022; Truckenbrodt et al., 2018). Most of these protocols have been developed for specimens that are at most 100 micrometers thick. We developed ExM methods for centimeter-scale tissue samples, including entire mouse brains. A key requirement for ExA-SPIM is optical clearing so that the entire volume can be imaged with diffraction-limited resolution without sectioning. Index of refraction inhomogeneities in heavily myelinated fiber tracts pose particular challenges. Clearing was achieved by stringent dehydration and delipidation prior to gelation and expansion. We systematically evaluated dehydration agents, including methanol, ethanol, and tetrahydrofuran (THF), followed by delipidation with commonly used protocols on 1-mm-thick brain slices. Slices were expanded and examined for clarity under a macroscope. Dehydration using THF was followed by two sequential delipidation steps. First DISCO type clearing Chi et al., 2018; Renier et al., 2014 followed by aqueous delipidation (Chen and Svoboda, 2020c, dx.doi.org/10.17504/protocols.io.zndf5a6) rendered the samples extremely transparent (Figure 3a; Ouellette et al., 2023, dx.doi.org/10.17504/protocols.io.n92ldpwjxl5b/v1). In addition to clearing specimens, delipidation facilitates immunolabeling brain samples prior to gelation and expansion (Figure 3). Signal amplification facilitates high contrast imaging of small structures, especially since the increase in volume upon expansion dilutes the concentration of the fluorophore. For gelation, we used VA-044 as the initiator, instead of the more commonly used APS/TEMED (Tillberg and Chen, 2019). VA-044 initiates free radicals at a temperature-dependent rate. At low temperatures (4 °C), the gelling reagents are allowed to diffuse to the center of thick samples, followed by higher temperatures (37 °C) to trigger uniform polymerization. Our protocol provides nearly isotropic expansion of the whole sample (including internal brain structures). Although developed for the mouse brain, we have successfully used this protocol for other large specimens, such as a 1 cm×1 cm×1.5 cm piece of macaque motor cortex, as well as a 1 cm×1 cm×0.01 cm section of human visual cortex (Ouellette et al., 2023, dx.doi.org/10.17504/protocols.io.n92ldpwjxl5b/v1).

We cleared and expanded entire mouse brains and imaged individual neurons, resolving their dense axonal projections (Figure 3). Tracking the axons of individual neurons is a challenging problem—axon collaterals can be very thin (<100 nm) and traverse large distances (centimeters), spanning vast areas of the brain. This requires high-resolution, high-contrast imaging of the entire brain without loss of data. Current best-in-class approaches (Economo et al., 2016; Gong et al., 2013; Gong et al., 2016; Winnubst et al., 2019) require physical sectioning and extensive tiling, which complicates downstream data processing. In addition, imaging lasts multiple days, which increases the chance for experimental failure and data loss. Finally, the resolution of these imaging methods is highly anisotropic (typically <1:6), which compromises the ability to perform unambiguous axon tracing.

We expressed GFP in a sparse subset of neurons in an enhancer virus mouse to label subcortical projection neurons. Brains were expanded (3×) and imaged in 15 tiles in as little as 24 hr. Because of the lack of physical tissue slicing and minimal tiling, datasets are aligned with an average residual error of <2 pixels between interest point correspondences in neighboring tiles (Hörl et al., 2019). Tile alignment of a representative dataset based on initial stage coordinates (Video 1 and Video 2), after stitching optimization (Video 3), and after fusion (Video 4) are available as videos. Dense axonal projections and varicosities are clearly visible (Figure 3e–h), and long-range axons can be tracked across the brain (Figure 3c–d). Videos of the data displayed in Figure 3e and f are shown in Video 5 Video 6.

Representative images of axons as a function of imaging depth are shown in (Figure 3—figure supplement 1). The achromatic performance of the ExA-SPIM optics also enables multiple fluorescence channels to be acquired from mouse brains (e.g. GFP and tdTomato; Figure 4).

We next established that our methods are applicable to larger brains. With minor adaptations, we applied our clearing and expansion protocol to a 1 cm×1 cm×1.5 cm block of pigtail macaque brain from the hand-wrist and trunk regions of primary motor cortex. Due to its strong myelination, this region of the primate brain is particularly difficult to clear and image (Van Essen et al., 2019). The specimen contained cortico-spinal neurons expressing a fluorescent protein (tdTomato) driven by retrograde AAV injected into the intermediate and ventral laminae of the cervical spinal cord. ExA-SPIM image volumes revealed brightly labeled cortico-spinal neurons (Figure 5a–b and Video 7, Video 8, Video 9). Individual neurons, their dendritic arbors, and extensive dendritic spines as well as the descending axon and collaterals are clearly discernible (Figure 5c–f and Video 10). Additionally, even in down-sampled data (~1 µm effective voxel size), we can follow axonal pathways. This raises the possibility that with a small number of slices (~6 × 1 cm thick slabs) and reduced imaging resolution, a modified inverted (Wu et al., 2013; Kumar et al., 2014) or open-top (McGorty et al., 2015; Glaser et al., 2017; Glaser et al., 2019; Barner et al., 2020; Glaser et al., 2022) ExA-SPIM design (Figure 5—figure supplement 1) could provide a mesoscale map of white matter axons across the entire macaque brain in several days. Existing efforts to map pathways in the primate brain at this resolution require slicing the brain in 300 µm sections and laboriously assembling the resulting images into a coherent 3D volume (Xu et al., 2021).

Heavy chain neurofilaments comprise the internal scaffolds of long-range projection axons. Visualizing these neurofilaments with immunofluorescence could provide detailed information about axonal trajectories without the need for gene transfer methods, and which cannot be achieved with other methods, such as diffusion magnetic resonance imaging (Huang et al., 2021; Van Essen et al., 2019) or high-resolution optical coherence tomography (Li et al., 2019). We next evaluated ExA-SPIM for imaging immunolabeled heavy chain neurofilaments in the human neocortex (Figure 6).

We cleared and expanded (4×) a~1 cm×1 cm×0.01 cm piece of human neocortex (Figure 6a) and labeled the tissue with fluorescent SMI-32, which preferentially stains heavy chain neurofilaments. Individual axons and their trajectories are clearly visible through the 350-µm-thick sample (Figure 6b–d). The larger axons (>1 µm) are well separated from each other in both the gray and white matter. In the white matter, axons are arranged as multiple intercalated populations coursing in different directions, rather than as homogeneous fascicles. These results demonstrate the feasibility of using ExA-SPIM for visualizing white matter tract axons and lay the foundation for scaling this data acquisition to multiple, thick tissue sections, and ultimately the entire human brain.

A parts list and CAD model are available at https://github.com/AllenNeuralDynamics/exa-spim-hardware (Allen Institute for Neural Dynamics, 2023b). A ZEMAX model is available at https://github.com/AllenNeuralDynamics/exa-spim-optics (Allen Institute for Neural Dynamics, 2023a). The system has two main optical paths (Figure 2—figure supplement 4). The first excitation path shapes and delivers the light sheet to the specimen. Laser light is provided by a custom-made laser combiner containing a 200 mW 405 nm laser (LBX-405–180-CSB-OE, Oxxius), 1000 mW 488 nm laser (Genesis MX488-1000 STM OPS Laser-Diode System, Coherent Inc), 1000 mW 561 nm laser (Genesis MX561-1000 STM OPS Laser-Diode System, Coherent Inc), and 1000 mW 639 nm laser (Genesis MX639-1000 STM OPS Laser-Diode System, Coherent Inc; Figure 2—figure supplement 5). An acousto-optic tunable filter (AOTFnC-400.650-TN, AA Opto Electronic) controlled by a RF driver (MPDS8C-D65-22-74.158-RS, AA Opto Electronic) is used to modulate each laser and control the output power of the combiner. The output beam from the combiner is ~1.2 mm in diameter, which is expanded to 13.3 mm using two relay lenses with f=3.6 mm (LMPLANFLN 50×, Olympus) and f=40 mm (HLB M PLAN APO 5×, Shibuya Optical). The light path is then reflected vertically by a third kinematic mirror and focused along one axis by an achromatic cylindrical lens (ACY254-050-A-ML, Thorlabs). The cylindrical lens is mounted in a motorized rotation mount (C60-3060-CMR-MO, Applied Scientific Instrumentation). This enables precise electronic control over the rotation of the light sheet within the specimen. Control over this parameter is important, as the field of view (~10.6 mm) and depth of field of the detection lens (<10 µm) require the sheet to be rotated with <1 deg precision. The light from the cylindrical lens is focused onto an electrically tunable lens (EL-16–40-TCVIS-20D-C, Optotune AG). The actuating surface of the electrically tunable lens is conjugated to the back focal plane of the excitation objective (1–290419, Navitar) through two large 20 mm galvanometric scanning mirrors (QS20X-AG, Thorlabs), three additional kinematic mirrors, and a final relay consisting of a first lens with f=200 mm (AC508-200-A-ML) and second lens with f=300 mm (AC508-300-A-ML). The two galvanometric mirrors are not conjugated to the back focal plane of the excitation objective and are used in tandem to tilt and translate the light sheet to be co-planar with the detection lenses’ focal plane. Similar to the cylindrical lens rotation, the light sheet must also be tilted to <1 deg precision across the ~8.0 mm vertical height of the imaging field of view.

The excitation objective has a focal length of f=110 mm, NA = 0.133, and back aperture diameter of ~29.26 mm, with which the preceding excitation optics results in a Gaussian excitation light sheet with NA ~0.1. The intensity profile across the width of the light sheet is also Gaussian, with a full-width half-maximum (FWHM) of ~12.5 mm. This corresponds to a reduced (60%) light intensity at the edges of the field of view. The lens is mounted in a custom dipping cap which extends the effective working distance from 39 mm in air to ~52 mm in water. The 1-mm-thick window on the dipping cap was sealed with UV curing optical adhesive (NOA86, Norland Products). The lens is oriented vertically, such that the light sheet is delivered downward into the immersion chamber which is filled with an immersion medium.

Fluorescence is collected with a high etendue electronics metrology lens (VEO_JM DIAMOND 5.0×/F1.3, Vieworks Co., LTD; jointly developed and fabricated by Schneider-Kreuznach). Although the spherical aberration introduced in the excitation path is negligible, the same dipping cap design would result in severe spherical aberration on the detection path, where the aperture was larger (NA = 0.305). Imaging at these higher apertures necessitated imaging lenses designed for direct immersion into a liquid mounting medium. The VEO_JM DIAMOND detection lens is designed for co-axial illumination, involving a 35-mm-thick BK7 glass (n=1.52) beam splitter mounted on the front object side of the lens. The beam splitter couples white light into the lens for bright illumination of electronic components during inspection. This was leveraged for aberration-free imaging into a liquid mounting medium.

We removed the beam splitter and replaced it with an optically equivalent thickness of the aqueous mounting medium (30–35 mm for refractive indices between 1.33–1.56) and glass (4 mm total, including the chamber window and fluorescence filter). This scheme avoids the spherical aberrations that would otherwise plague imaging with an air-immersion lens into a non-air-based medium. The lens is positioned horizontally on the optical table outside of the immersion chamber. The lens images the specimen through a λ/10 50 mm diameter VIS-EXT coated fused silica window that is 3 mm thick (#13–344, Edmund Optics). The window was glued to the immersion chamber using UV curing optical adhesive (NOA86, Norland Products). A custom 44 mm diameter 1-mm-thick multi-bandpass fluorescence filter (ZET405/488/561/640mv2, Chroma) is inserted on the outer side of the immersion chamber and held in place with a retaining ring (SM45RR, Thorlabs). Placing the filter on the object side of the lens results in a cone angle of incidence of 17.8 deg. This results in a spectral shift and broadening of transmitted light, which was modeled to account for adequate suppression of scattered light at the excitation laser wavelengths.

The lens is attached to the camera (VP-151MX-M6H00, Vieworks Co., LTD) using a custom mechanical assembly that enables tipping and tilting of the lens and camera relative to the immersion chamber window. To reduce fixed pattern noise at low light levels, we requested that the manufacturer permanently disable the photo response non-uniformity (PRNU) correction mode. The camera uses the Sony IMX411 sensor, which features a 14192×10,640 array of 3.76 µm pixels. The ×5.0 magnification of the detection lens implies sampling at 0.75 µm in the specimen plane. Because the detection lens is not infinity-corrected, it is not possible to change tube lenses to change the pixel sampling. Therefore, although the optical resolution of the detection lens is ~1.0 µm, the sampling limited resolution based on the Nyquist criterion is 1.5 µm. However, it is worth noting that Vieworks Co. produces cameras with pixel-shifting technology, which would enable finer sampling if needed. The cooling fan on the camera was replaced with a quieter fan operating at 2400 rpm (Noctua NF-A6x25 FLX, Premium Quiet Fan, 3-Pin, 60 mm, Noctua). Using an alignment laser, the entire assembly was aligned to the window. The entire assembly is also mounted on a rail system (XT95, Thorlabs) that enables precise axial alignment of the assembly relative to the chamber window.

The sensor can be operated with 12, 14, or 16-bit analog to digital (A/D) conversion, each of which provides a trade-off between noise and data throughput. When operated with 14-bit A/D, the line time of the sensor is 20.15 µs. With 14192 pixels per row on the sensor, this corresponds to an imaging speed of 703×10⁶ voxels/s. The line times and imaging speeds at 12 and 16-bit are 15.00 and 45.44 µs respectively, corresponding to 946 and 312×10⁶ voxels/s. This contrasts with a state-of-the-art sCMOS sensor with 2048 pixels per row, where even at the fastest 4.89 µs line time, the imaging speed is only 418×10⁶ voxels/s. In other words, the Sony IMX411 sensor provides more pixel parallelization within each row. This enabled twice the voxel rate with four times the pixel dwell time. See Appendix 1 for further discussion on comparing the speed and sensitivity of cameras.

The specimens, in this case expanded gels, were mounted in a customized holder, which was attached to a motorized XY stage (MS-8000, Applied Scientific Instrumentation) and Z stage (LS-100, Applied Scientific Instrumentation), for scanning and tiling-based image acquisition.

The microscope was controlled using a high-end desktop workstation (SX8000, Colfax Intl). The workstation was equipped with a motherboard with five PCIe 4.0x16 and one PCIe 4.0x8 slots (X12DAI-N6, Supermicro), most of which are required for the various electronics needed to control the ExA-SPIM. One slot was used for the frame grabber (Coaxlink Octo, Euresys), which streams imaging data from the camera onto a fast local 12.8 TB NVME drive (7450 Max 12800 GB 3 DWPD Gen4 15 mm U.3, Micron). A second slot was used for the data acquisition (DAQ) card (PCIe-6738, National Instruments) used for generating the various digital and analog voltage signals. A third slot was used for the high-speed network interface card (ConnectX-5 EN MCX515A-CCAT QSFP28 Single Port 100GbE, Mellanox) to transfer data off the local NVME drive, over the network, and onto a networked server. A fourth slot was occupied by the workstation GPU (A4000, NVIDIA). The workstation was also connected to a controller (TG-16, Applied Scientific Instrumentation) via a USB connection. The TG-16 controller was equipped with cards for controlling the motorized X, Y, and Z stages, the electrically tunable lens, and the cylindrical lens rotation mount. Finally, each laser within the combiner was connected to the computer via a USB connection, as well as the RF driver of the AOTF. The DAQ acted as the master controller of the entire system. Analog output voltages were used to drive or trigger the electrically tunable lens, two scanning galvanometric mirrors, camera, scanning stage, and the RF outputs to the AOTF for each laser.

The microscope was controlled using custom software written in Python https://github.com/AllenNeuralDynamics/exaspim-control (Allen Institute for Neural Dynamics, 2025b) and was operated on the Windows 10 operating system. The software could control and configure all of the electronically controlled hardware devices and uses Napari https://napari.org as the graphical user interface (GUI) for image streaming and visualization. An imaging experiment consisted of a series of nested loops, which included looping over the total number of frames within a given tile, looping over the total number of channels within a tile, and finally tiling in two dimensions to cover the entire tissue volume. A multiprocessing, double buffering scheme was used to capture a tile while the previous tile was being transferred over the network to longer term storage. All hardware was controlled using a custom library developed for generalized microscope control https://github.com/AllenNeuralDynamics/voxel (Allen Institute for Neural Dynamics, 2025a).

To achieve robust and reliable acquisition at the speed required by the ExA-SPIM, we used two data streaming strategies. The first used the open-source eGrabber frame grabber Python API and open-source Acquire package (https://github.com/acquire-project) (Clack et al., 2023), specifically the Python API Acquire-Zarr (Liddell et al., 2025). Acquire is a new state-of-the-art microscope acquisition project which enables the ExA-SPIM system to stream data at the required rate, directly to OME-Zarr file format (V2 or V3), with either ZStandard or LZ4 compression, a variable chunk size, a variable shard size (for V3), and an optional multi-resolution pyramid, all of which help streamline downstream data storage and handling. The second strategy used the ImarisWriter library (Beati et al., 2020). This option enabled high-speed data streaming with online lossless compression (LZ4 with bit shuffling) and real-time writing of a multi-resolution pyramid.

On average, the lossless compression ratio using either ImarisWriter or Acquire was ~1.5–4×depending on the bit mode of the camera (e.g. 12, 14, or 16 bits). This reduced the overall effective data rate of the system and eased network transfers of the data to centralized storage. A schematic of the acquisition pipeline is shown in Figure 2—figure supplement 9.

To quantify the point spread function of the ExA-SPIM microscope, we imaged fluorescent 0.2 µm TetraSpeck microspheres (Invitrogen, cat no. T7280, lot no. 2427083) using 561 nm excitation. A cube of expanding hydrogel containing a 10% (v/v) microbead solution was prepared, and 40 µL of microspheres solution was added to 360 µL of activated Stock X monomer solution, as described in Asano et al., 2018. The solution was briefly vortexed and carefully pipetted into an array of 2 mm³ wells in a silicone mold. The mold was placed in a sealed petri dish containing damp Kimwipes to maintain humidity and incubated at 37 °C for 2 hr. After incubation, the polymerized bead phantoms were placed into 0.05×SSC buffer for at least 24 hr to expand and equilibrate prior to imaging. After volumetric imaging, the resulting imaging stack was analyzed using custom-written Python code https://github.com/AllenNeuralDynamics/exa-spim-characterization (copy archived at Allen Institute for Neural Dynamics, 2023c). The results shown in Figure 2 are averaged from >10,000 individual beads.

We used the field curvature quantification methods developed for the Benchtop mesoSPIM (Vladimirov et al., 2024). Briefly, the field curvature was measured using a high-precision Ronchi ruling with 120 lines per mm (62–201, Edmund Optics). The ruling was mounted into the system and aligned to be normal (i.e. flat) with respect to the imaging path. The ruling was trans-illuminated using light-emitting diodes (LEDs) at 450 nm (M450LP2, Thorlabs), 530 nm (M530L4, Thorlabs), 595 nm (M595L4, Thorlabs), and 660 nm (M660L4, Thorlabs). To provide uniform illumination, the LEDs were collimated to a diameter of ~20 mm and passed through a diffuser (ED1-C20, Thorlabs). Image stacks were captured by scanning the Ronchi ruling in 1 µm steps through the imaging path focal plane (2 mm total scan range) with illumination at either 450 nm, 530 nm, 595 nm, or 660 nm. The resulting image stack was split into a 16×16 grid of regions of interest (ROI), each 887×665 × 2000 pixels. Within each ROI, the contrast of each frame in the stack was calculated using the 5th and 95th percentiles of intensity within the frame, where the contrast, C = (I_max – I_min)/(I_max +I_min). The resulting contrast versus depth curve was fit to a normal distribution to extract the depth (i.e. index) corresponding to maximum contrast. The indices were radially averaged across the 16×16 grid of ROIs, yielding an estimate of the imaging lens field curvature within each of the four tested wavebands. Custom Python scripts were used to run the analysis https://github.com/AllenNeuralDynamics/exa-spim-characterization.

The lens distortion was measured using a target (62–950, Edmund Optics) with 125 µm diameter dots spaced every 250 µm in a grid pattern. The target was mounted and trans-illuminated using LEDs in the same manner as the field curvature quantification. A single image was acquired with the target at the focal plane of the imaging lens. The resulting image was quantified by first segmenting and calculating the centroid of each dot. The calculated position of each dot was then compared to the theoretical dot position. The percent distortion for each dot was defined as the difference between the experimental and theoretical positions, normalized by the theoretical position. The resulting distortion values were radially averaged, yielding the lens distortion as a function of position from the center of the lens field of view. Custom Python scripts were used to run the analysis https://github.com/AllenNeuralDynamics/exa-spim-characterization.

A bead phantom (see point spread function measurement) was used to compare the sensitivity of the large-format CMOS camera and a scientific CMOS camera (Orca Flash V3, Hamamatsu). A single bead was first imaged with the scientific CMOS camera, varying the effective pixel rate or imaging speed. The scientific CMOS camera was then removed from the microscope and replaced with the large-format CMOS camera. The same bead was located and imaged again at various effective pixel rates. The bead was located in all image stacks, and the signal-to-noise ratio of the bead was quantified as the signal of the bead, minus the average background signal, divided by the background noise. A custom written Python code was used to run the analysis https://github.com/AllenNeuralDynamics/exa-spim-characterization . It is worth noting that the Orca Flash V3 is an older scientific CMOS camera. Newer cameras, such as the Orca BT Fusion with improved noise characteristics (1.0 e- versus 1.6 e- read noise and 95% vs 85% peak quantum efficiency), are now available. However, these differences would not significantly change the results shown in Figure 2e. Appendix 1 contains a deeper discussion on comparing the speed and sensitivity of cameras.

Data was streamed using the Acquire or ImarisWriter onto a local NVME drive in the acquisition workstation. Upon the completion of a tile, the resulting files were transferred over the high-speed network using xcopy onto a 670 TB enterprise storage server from VAST data. This occurred in parallel with the acquisition of the next imaging tile. Upon completion of the transfer, the tile was deleted off the acquisition workstation’s NVME drive.

The raw data for each color channel consisted of a set of overlapping 3D image tiles. For the ImarisWriter acquisition workflow, the images were stored in Imaris IMS file format, an HDF5-based format which enables fast parallel writing to local storage. However, this format is not well suited to being archived in cloud (object) storage since accessing an arbitrary data chunk requires seeking a file. Thus, the data was converted to OME-Zarr format, which supports parallel read-write over the network and has a rich metadata structure (Moore et al., 2021; Moore et al., 2023). OME-Zarr is also integrated with visualization and annotation tools used for downstream visualization and analysis, including Horta Cloud (https://github.com/JaneliaSciComp/hortacloud; Clements et al., 2025) and Neuroglancer (https://github.com/google/neuroglancer; Maitin-Shepard, 2025; Maitin-Shepard et al., 2021). With up to four channels, the storage footprint for a single dataset can reach hundreds of terabytes. Therefore, the storage ratios and compression/decompression speeds of various lossless codecs were compared. Blosc ZStandard yielded the highest storage ratios, with compression speeds comparable to LZ4 at lower ‘clevel’ settings (e.g. 1–3). Dask was used to parallelize the compression and OME-Zarr writing over a high-performance computing (HPC) cluster consisting of 16 nodes, each with 32 Intel CPUs with Advanced Vector Instructions 2 (AVX2) and 256 GB RAM. Image chunks were read in parallel from high-bandwidth network storage, compressed in memory, and written directly to AWS S3 and Google Cloud Storage buckets. Combined throughput (chunk read, compress, write) reached over 2 GB/s, with execution time dominated by read-write I/O operations. The image compression and cloud transfer code is available at: https://github.com/AllenNeuralDynamics/aind-data-transfer (Arshadi, 2024). For the Acquire acquisition workflow, datasets were streamed directly to the OME-Zarr format.

ExA-SPIM image stitching used a combination of on-premises and cloud-based resources. Datasets were first converted from OME-Zarr to N5 and stitched using BigStitcher on the Google Cloud Platform (GCP) (Hörl et al., 2019). Tile placement transformations were based on interest points. The interest point-based tile registration consisted of three steps: identification of interest points, finding corresponding interest points between tiles, and optimizing for tile transformation parameters with respect to the distance of corresponding interest points. We first performed a translation-only stitching routine, followed by a full affine transformation optimization. The affine transformation was regularized by a rigid transformation, and corner tiles were kept fixed to prevent the global scaling of the sample and divergent solutions at the corner tiles. As a rule of thumb, several thousand interest points per overlapping region were required for a reliable identification of correspondences and successful optimization. Once the alignment transformations were calculated, the tiles were fused using the on-premises HPC into a single contiguous volume in the N5 format. A representative dataset (~100 TB in raw uncompressed size) can be fused in ~24 hr using 16 nodes (32 cores each, 480 total), with 16 GB of RAM per core (7.68 TB total) and an output chunk size of 256, 256, 256 pixels. 2 cores per node are reserved as overhead per spark worker (i.e. 32 total for 16 nodes). Fused datasets along with the raw tiled datasets and the tile placement transformations are deposited in an Amazon Web Services S3 aind-open-data data bucket. The N5 datasets were converted to OME-Zarr for visualization with Neuroglancer, or KTX (https://registry.khronos.org/KTX) for visualization with HortaCloud. The current pipeline involves several file format conversion steps and transfers between on-premises and cloud-based platforms. Future pipelines will standardize around OME-Zarr and be completely operable in the cloud. Only the initial data conversion and compression step would be computed on-premises.

To generate whole-brain single neuron reconstructions, we use HortaCloud, an open-source streaming 3D annotation platform enabling fast visualization and collaborative proofreading of terabyte-scale image volumes. Human annotators proofread stitched image volumes using HortaCloud in a web browser on a personal workstation. Starting from the soma, the axonal and dendritic arbors were traced through all terminals by laying down connected points along the neurite, producing a piecewise-linear approximation of neuronal trees (Economo et al., 2016).

Briefly, the samples were cleared, immunolabeled, and expanded as described below. Detailed protocols for preparing cleared and expanded brains are available at [(Ouellette et al., 2023), dx.doi.org/10.17504/protocols.io.n92ldpwjxl5b/v1].

The process for preparing the macaque brain samples is described below.

Deidentified postmortem adult human brain tissue (61-year-old male, Hispanic, no known history of neuropsychiatric or neurological conditions) was obtained with permission from next-of-kin by the San Diego Medical Examiner’s Office. Tissue procurement was reviewed by the Western Institutional Review Board (WIRB) and did not constitute human subject research requiring Institutional Review Board (IRB) review, in accordance with federal regulation 45 CFR 46 and associated guidance. Postmortem tissue collection was performed in accordance with the Uniform Anatomical Gift Act described in Health and Safety Code §§ 7150, et seq., and other applicable state and federal laws and regulations.

Tissue was manually sliced into 1 cm coronal slabs, flash frozen with liquid nitrogen, vacuum sealed, and stored at –80 °C by the Allen Institute Tissue Processing Team. One slab from the occipital pole was drop-fixed in 4% PFA for approximately 12 hr at 4 °C, and tissue regions containing visual cortex were dissected into ~1 cm blocks for histological processing. Individual blocks were then SHIELD-fixed (Lifecanvas Technologies) prior to sectioning at 100 µm on a sliding freezing microtome to further protect protein antigenicity and tissue architecture. Individual free-floating sections were then passively delipidated (LifeCanvas Technologies) for 1 week prior to immunolabeling. Delipidated sections were subsequently immunolabeled for SMI-32, which identifies heavy chain neurofilaments that make up axon scaffolds in long-range projection neurons. Free-floating sections were blocked in NGSTU (5% goat serum, 0.6% Triton X-100, 4 M urea in 1×PBS) overnight, incubated with a primary antibody (rabbit anti-neurofilament 200, Sigma Aldrich N4142) diluted 1:500 in NGSTU + 0.02% sodium azide for 5 days, followed by a secondary antibody incubation (goat anti-rabbit AF488, Thermo Fisher A-11034) diluted 1:100 in NGST (5% goat serum, 0.6% Triton X-100 in 1×PBS) for 4 days. Due to the thinness (100 um) of this tissue compared to an entire mouse brain, the gelling protocol used varied slightly from the whole brain gelling protocol described above. Most notably, the thermal initiator used was ammonium persulfate (APS) instead of VA-044, the tissue was polymerized at room temperature for 3 days, and the resulting tissue-hydrogel matrix was digested using a 1:50 concentration of proteinase-k in buffer (5% Sodium Dodecyl Sulfate, 5% Triton X-100, 10% 1 M TRIS pH8).

The expanded samples were trimmed to produce smooth edges and then were placed in a custom-built anodized imaging chamber. The sample chamber assembly was performed in a large bath of the expansion and imaging solution (0.05×SSC). The smooth edges of the hydrogel were placed against the chamber panels corresponding to the excitation and emission path, and then the chamber was removed from the bath. A warm (55 °C) solution of 2% agarose, in the same 0.05×SSC as used during expansion, was carefully poured into the chamber space behind the hydrogel for structural rigidity during imaging and let cool to room temperature until solid. The chamber was then sealed and placed in 0.05×SSC for equilibration overnight before imaging. The detailed protocol for mounting the expanded hydrogels is available on protocols.io: (Ouellette et al., 2023), https://www.protocols.io/view/whole-mouse-brain-delipidation-immunolabeling-and-cmqbu5sn?step=12.

Code is available as listed below. Please see the Materials and methods section for additional usage details. Microscope control: https://github.com/AllenNeuralDynamics/exaspim-control (Allen Institute for Neural Dynamics, 2025b) Acquire video streaming: https://github.com/acquire-project Hardware files: https://github.com/AllenNeuralDynamics/exa-spim-hardware (Allen Institute for Neural Dynamics, 2023b) Optical files: https://github.com/AllenNeuralDynamics/exa-spim-optics (Allen Institute for Neural Dynamics, 2023a) Characterization scripts: https://github.com/AllenNeuralDynamics/exa-spim-characterization (copy archived at Allen Institute for Neural Dynamics, 2023c) Compression and cloud transfer: https://github.com/AllenNeuralDynamics/aind-data-transfer.