Methods, Trade-offs, and Opportunities in High-speed Optical Microscopy for Neural Voltage imaging

Abstract

Electrical signals in neurons underlie perception, movement, memory, and behaviour, yet many unfold too rapidly to be captured by conventional optical imaging. Calcium imaging has transformed neuroscience but provides an indirect and relatively slow readout of electrical activity. By directly measuring membrane-potential changes, voltage imaging enables millisecond-scale recording of action potentials, subthreshold dynamics, and signal propagation across neural circuits. Recent advances in voltage-sensitive dyes and genetically encoded voltage indicators have made voltage imaging increasingly practical, motivating the development of fluorescence microscopy methods optimized for high-speed acquisition. However, voltage imaging remains constrained by trade-offs among imaging speed, spatial resolution, signal-to-noise ratio, and photodamage. In this Review, we discuss high-speed optical microscopy strategies that address these challenges and highlight the need for co-design among voltage indicators, imaging systems, and computational analysis.

INTRODUCTION

Communication between neurons is driven by rapid membrane-potential fluctuations, so understanding neural circuits requires tools that can record activity across large neuronal populations in vivo. Optical imaging complements electrode-based methods by enabling minimally invasive, simultaneous recordings from many neurons, spatial mapping of activity and access to genetically defined cell classes. It can also capture subthreshold dynamics, monitor neuronal subcompartments, and reveal spatiotemporal patterns, such as signal propagation across circuits^1–6, that are difficult to resolve with conventional electrophysiology.

Among optical approaches, calcium imaging has become a widely used workhorse for monitoring population neural activity and has enabled numerous foundational discoveries in neuroscience.^7–9 However, calcium signals provide an indirect and relatively slow proxy for electrical activity^10,11, reflecting changes in intracellular calcium rather than membrane potential itself. As a result, calcium imaging is poorly suited for resolving precise spike timing, subthreshold events such as excitatory and inhibitory postsynaptic potentials, or hyperpolarizing signals, rendering key aspects of synaptic integration, inhibition, and plasticity inaccessible.^12–14 By contrast, voltage imaging directly reports membrane potential with millisecond temporal precision, providing a more faithful view of neuronal activity.^1,3,15–17

Despite this promise, voltage imaging imposes stringent technical demands. During an action potential, the electric field reorients within one to two milliseconds, with the fastest component—the rising phase—occurring in approximately 250 μs (Box 1). This makes action potentials 100–1000-fold faster than calcium transients, necessitating high imaging speeds to accurately capture their dynamics.¹⁸ In practice, voltage imaging is typically performed at frame rates over 300 Hz¹⁴ and preferably in the kilohertz range^18,19 to avoid waveform distortion or missed events.

Box 1 |. Membrane Action Potential Signalling in Neurons.

Neurons integrate synaptic inputs across their dendrites and soma. When membrane depolarization reaches threshold, an action potential is initiated by the rapid activation of voltage-gated Na⁺ channels and subsequently repolarized by voltage-gated K⁺ channels. This electrical signal propagates along the axon and mediates communication with downstream neurons through chemical neurotransmission or electrical coupling. Modern voltage indicators directly report membrane potential by converting voltage-dependent changes, such as protein conformational shifts or electrochromic effects, into fluorescence signals. As an example, in the ASAP family of GEVIs, a fluorescent protein (FP) is coupled to a voltage-sensing domain (VSD), such that voltage-driven conformational changes in the VSD alter fluorescence intensity.

Such acquisition rates severely limit photon collection time, making voltage imaging intrinsically noisier than calcium imaging. Increasing excitation intensity can partially compensate for reduced exposure time, but at the cost of faster photobleaching, shorter recording durations, and increased risk of photodamage due to heating or nonlinear effects. In addition, voltage signals originate from the plasma membrane, where the number of contributing fluorophores is inherently limited, further constraining signal strength. Finally, imperfect membrane targeting of voltage indicators can increase background fluorescence from non-responsive molecules, elevating shot noise, and degrading both the signal-to-background ratio (SBR) and signal-to-noise ratio (SNR).

On the detection side, maintaining a large, or three-dimensional (3D), field of view (FOV) while resolving fine neuronal features remains challenging because kilohertz imaging demands immense pixel throughput. Sustaining these data rates over extended recordings further stresses data transmission bandwidth, real-time throughput, and storage capacity. In this review, we highlight recent advances in optical microscopy that help overcome these constraints and enable high-speed one-photon (1P) and two-photon (2P) voltage imaging. Collectively, these developments expand our ability to capture rapid neuronal dynamics with high spatiotemporal fidelity, opening new opportunities to probe cooperative signalling and large-scale organization in neural networks.

2. STATE-OF-THE-ART ONE-PHOTON VOLTAGE IMAGING

Currently, 1P excitation microscopy remains the dominant platform for voltage imaging because it provides efficient excitation and supports high-throughput widefield detection. Here, we review state-of-the-art 1P methods in three categories: planar approaches for high temporal throughput, multi-depth strategies for optical sectioning and volumetric access, and snapshot light-field methods for camera-limited 3D imaging. Representative performance metrics are summarized in Table 1.

Table 1. Comparison of representative 1P/2P voltage imaging modalities in vivo.

For techniques demonstrated under multiple experimental configurations, we report each operating regime as a separate sub-entry with matched FOV, acquisition rate, and other parameters. The “# Neurons (best reported)” column indicates the maximum number of neurons simultaneously recorded in a single imaging session. The “continuous recording duration (best reported)” refers to the longest uninterrupted recording duration reported by the original study.

Technique (Year)	FOV (2D: diameter or X×Y; 3D: diameter × thickness or X×Y×Z)	Frame/Volume rate (Hz)	Continuous recording duration (best reported)	# Neurons (best reported)	Voltage indicator	Model system applied
1P
MiniVolt104 (2026)	∅250 μm (2D)	530	—	6	Voltron2	Mouse visual cortex V1(L1)
SLIM51 (2025)	∅550×300 μm3 (3D)	800	180 s	27	pAce	Mouse hippocampal CA1(SP)
Confocal LFM54 (2024)	∅800×180 μm3 (3D)	400	40 s	330	Voltron	Mouse visual cortex V1(L1 and upper portion of L2/3)
FLIPR113 (2024)	390×46×50 μm3 (3D)	500	20 s	>100	Voltron2	Zebrafish larvae spinal cord (3 to 4 days post-fertilization).
TICO30 (2024)	1,160 × 325 μm2 (2D)	800	20 min	78	Voltron2 & somArchon	Mouse neocortex (L2) and hippocampal CA1(SP)
MuZIC33 (2023)	150×150×45 μm3 (3D)	916	—	40	Voltron2-ST	Mouse neocortex (L2/3)
Wang, Zeguan, et al., bioRxiv (2023)	930×370×170 μm3 (3D)	200.8	35 s	25,556	Positron2-Kv	Zebrafish larvae whole brain (5 to 6 days post-fertilization)
VsLFM53 (2023)	260×260×100 μm3 (3D)	500	5 s	—	pAce	Adult Drosophila whole brain
2P
FRAME (Liu, Zheng, et al., bioRxiv) (2025)	96×120 μm2	1006	—	Single	JEDI-2P-Kv	Mouse visual cortex V1 (L2/3)
	320×120 μm2	805	30 s	>200	JEDI-2P-Kv	Mouse hippocampal CA1 (SP)
FACED 275 (2025)	160×400 μm2	800	8.25 s	182	JEDI-2P-Kv	Mouse visual cortex V1 (L2/3)
	320×400 μm2	385	8.32 s	225	JEDI-2P-Kv	Mouse visual cortex V1 (L2/3)
AES77 (2024)	31×365 μm2	1648	—	2	ASAP5	Mouse visual cortex V1 (L5/6)
	78×365 μm2	701	—	≥2	ASAP5	Mouse visual cortex V1 (L5/6)
Scanless 2P78 (2024)	86×300 μm2	500	30 s	15	JEDI-2P-Kv	Mouse barrel cortex (L2/3)
SMURF76 (2023)	400×400 μm2	803	Near-continuous for 60 min (59s on, 1s off)	>100	SpikeyGi2	Mouse somatosensory cortex (L2/3)
ULoVE57 (2022)	—	440 (resonant scan)	>30 min	4	JEDI-2P-Kv	Mouse visual cortex V1 (L2/3)
	—	2500–5000 (random access)	>10 min	2	JEDI-2P-Kv	Mouse visual cortex V1 (L2/3, L5)
FACED74 (2020)	50×250 μm2	1000	1 s	15	ASAP3-Kv	Mouse visual cortex V1 (L2/3)

2.1. PLANAR VOLTAGE IMAGING

Conventional 1P voltage imaging often uses widefield microscopy, which captures a two-dimensional (2D) FOV in a single snapshot. This parallel detection provides a simple, low-cost route to the high frame rates needed for fast voltage dynamics. A key limitation, however, is the lack of optical sectioning: out-of-focus fluorescence contributes substantially to the detected signal, elevating background and reducing the SBR (Box 2), particularly in vivo. To address this challenge, recent work has combined widefield imaging with molecular strategies, including soma targeting^1,20–22 and sparse labeling^1,23, to reduce diffuse background and crosstalk. Complementary optical approaches further suppress background by restricting excitation to relevant structures, using structured^24,25, holographic^26,27, or targeted illumination^1,3,16,28,29.

Box 2 |. Out-of-focus fluorescence versus light scattering.

High-fidelity voltage imaging requires overcoming two distinct optical challenges. Their relative impact depends heavily on the biological preparation and dictates fundamental instrument trade-offs.

• Out-of-focus fluorescence arises when fluorophores outside the target focal plane are excited. In densely labeled tissue, this creates a diffuse background that consumes detector dynamic range and degrades the signal-to-background ratio, even in perfectly transparent samples. Light-field microscopy (LFM), for example, lacks intrinsic optical sectioning and is highly vulnerable to this background. To mitigate this, techniques like confocal microscopy and confocal LFM use physical apertures to reject out-of-focus emission, while light-sheet fluorescence microscopy avoids generating it via selective planar excitation. Two-photon (2P) microscopy minimizes it through spatially confined, nonlinear excitation. Alternatively, out-of-focus background can also be suppressed through sparse expression and labeling if the biological question allows.

• Light scattering occurs when photons are deflected by refractive index inhomogeneities. Scattering limits excitation penetration and causes scattered emission to form broad, overlapping spatial footprints, making it difficult to unambiguously demix the activity of individual neurons. It is the dominant barrier for deep in vivo imaging in mammalian brains, though negligible in transparent organisms like embryonic zebrafish. One-photon techniques rely primarily on ballistic photons, severely limiting depth. Conversely, 2P microscopy utilizes less-scattered near-infrared wavelengths and preserves spatial resolution at depth by assigning scattered emission photons to the scanning focal spot.

• Balancing trade-offs: Different architectures balance these physical constraints against kilohertz temporal demands. For example, while confocal LFM rejects background, it remains depth-limited by scattering. 2P microscopy elegantly solves both scattering and background challenges but requires complex spatial multiplexing or holographic strategies to achieve necessary speeds without inducing excessive photothermal damage.

Despite these advances, widefield microscopy has so far been used primarily at shallow depths and in sparsely labelled preparations. Imaging densely labelled populations and deeper brain regions instead favours optical-sectioning modalities that physically suppress out-of-focus fluorescence. However, conventional optical-sectioning approaches such as confocal microscopy typically rely on point scanning, which intrinsically limits frame rate, particularly over large FOVs, making it challenging to capture fast voltage dynamics across large neuronal populations.

To increase throughput, line-scan confocal microscopy has been adopted for voltage imaging, reducing the scanning requirement to one dimension by illuminating the sample with a focused line and detecting the resulting fluorescence with a linear detector array.³⁰ Although line scanning sacrifices some background rejection relative to point-scanned confocal microscopy, it enables kilohertz imaging over millimetre-scale FOVs. In addition, the reduced demands on scan speed permit the use of large-aperture scanners, facilitating efficient fluorescence collection at high numerical aperture (NA). To offset the diminished background suppression, Xiao et al. combined line-scan confocal microscopy with targeted illumination to selectively excite only neurons of interest, achieving more than a 50-fold improvement in SBR (Fig. 1a).³⁰

Figure 1. High-speed confocal and light-sheet 1P voltage imaging.

Confocal and light-sheet microscopy improve 1P voltage imaging by suppressing out-of-focus background while maintaining high acquisition speed. a, Line-scanning confocal detection combined with targeted illumination selectively excites cells of interest, reducing optical crosstalk and improving SBR for in vivo voltage recordings in mouse neocortex. b, Multi-plane confocal microscopy extends confocal voltage imaging to three dimensions by recording multiple axial planes simultaneously in vivo. c, Light-sheet microscopy uses selective planar illumination for optical sectioning and kilohertz-rate voltage imaging, shown here during motor activity recording in larval zebrafish spinal cord. d, Remote focusing enables rapid volumetric light-sheet imaging by shifting the focal plane with a fast-moving mirror, avoiding mechanical translation of the primary objective. DMD, digital micromirror device; PD, photodiode. Adapted with permission from refs.^30,33,37,41.

Wide-field fluorescence lifetime imaging also represents a promising new direction for planar voltage imaging. Using electro-optic fluorescence lifetime imaging microscopy (EO-FLIM), Bowman et al. achieved kilohertz-rate wide-field lifetime imaging of neuronal voltage signals.³¹ Because lifetime readout is less sensitive to photobleaching, this work highlights fluorescence lifetime as a potentially valuable complement to conventional intensity-based approaches.

2.2. MULTI-DEPTH VOLTAGE IMAGING

To extend voltage imaging to 3D, Multi-Z confocal microscopy³² employs axially extended illumination (via a lower NA) and a cascade of reflective pinholes in the detection path to acquire signals from multiple depths simultaneously (Fig. 1b). Fluorescence rejected by one pinhole is relayed to subsequent pinholes, minimizing signal loss and maintaining high overall detection efficiency. A recent variant, multi-Z imaging with confocal detection (MuZIC)³³, incorporates an ultrafast polygon scanner to achieve a line rate approximately an order of magnitude faster than resonant galvanometric scanners. Combined with an orthogonal galvo and axially distributed pinholes, MuZIC allows simultaneous scan of four imaging planes.²⁸ Further acceleration of Multi-Z confocal imaging can be achieved by adopting line scanning rather than point scanning.³⁴ Compared with conventional 2D imaging, multiplane imaging is more robust to axial brain motion because adjacent planes help distinguish true voltage signals from motion-induced artefacts, which is critical for in vivo imaging.

Additional gains in imaging speed can be achieved by further reducing scan dimensionality. Instead of scanning a point or a line, light-sheet fluorescence microscopy (LSFM) acquires images plane-by-plane: a thin sheet of light selectively illuminates a single optical section, and a 2D camera records that plane in a single exposure (Fig. 1c). This selective planar excitation provides optical sectioning while minimizing phototoxicity.^35,36,37

High-resolution voltage imaging with LSFM typically requires a high-NA detection objective, which comes with a shallow depth of field. Consequently, during 3D imaging the detection focal plane must be translated in synchrony with the axially scanned light sheet to maintain focus on the illuminated section. A common implementation physically moves the imaging objective using a piezoelectric actuator; however, the inertia of the objective limits the speed and efficiency of multi-plane acquisition. Remote focusing offers an attractive alternative for fast volumetric imaging: rather than translating the primary objective, an additional refocusing element is inserted into the detection path to rapidly shift the focal plane. This can be achieved by introducing controlled defocus using a tuneable lens in the Fourier plane or a piezo-driven mirror in a conjugate image plane.^38–40 The mirror-based approach is often preferred, as it can provide aberration-free refocusing over a relatively large axial range (Wang, Zeguan, et al., bioRxiv, 2023).

Mirror-based remote-focusing LSFM requires the incident rays to be nominally normal to the refocusing mirror to introduce controlled defocus. This is commonly implemented using a beamsplitter to co-align the incident and reflected paths, but the beamsplitter incurs substantial light loss—an acute limitation for voltage imaging, where exposure times are short and SNR is paramount. To overcome this inefficiency, Böhm et al. introduced flipped image remote focusing (FLIPR), which replaces the beamsplitter with a microscopic retroreflector (Fig. 1d).⁴¹ The retroreflector flips and folds the image back into the remote-focusing arm, approximately doubling light efficiency relative to conventional beamsplitter-based designs that use a polarization beamsplitter together with a quarter-wave plate.

The volumetric frame rate of remote-focusing LSFM is ultimately limited by the focal switching speed of the refocusing elements. To circumvent this constraint, Sacconi et al. introduced a parallelized detection strategy that uses multiple cameras to acquire seven planes simultaneously.⁴² In their implementation, several light sheets are projected into the sample in parallel, and fluorescence from each illuminated plane is directed to a dedicated camera. However, scaling camera-array approaches remains challenging: repeated beam splitting reduces the photon budget per plane (and thus SNR), and exciting multiple planes simultaneously increases out-of-focus background and degrades image contrast.

2.3. SNAPSHOT VOLUMETRIC VOLTAGE IMAGING

Light-field microscopy (LFM) is a leading snapshot 3D imaging modality and has been successfully applied to voltage imaging. Unlike conventional widefield microscopy, LFM captures both spatial and angular information using an angular-demultiplexing element such as a lenslet array (Fig. 2, central dashed circle). The resulting spatial-angular dataset can be computationally refocused to reconstruct volumetric images by leveraging disparities among angular views.^43–46 Importantly, because the full light field is recorded, optical aberrations can be estimated and corrected in postprocessing by digitally manipulating the captured rays (digital adaptive optics), eliminating the need for deformable mirrors or spatial light modulators.^47–49

Figure 2. Snapshot 3D light-field 1P voltage imaging.

Light-field microscopy (LFM) enables scanless volumetric imaging by capturing spatial and angular fluorescence information in a single camera snapshot. The central dashed region illustrates unfocused and Fourier LFM architectures, in which a lenslet array separates emitted light into perspective views for computational 3D reconstruction. a, Squeezed light-field microscopy (SLIM) compresses redundant perspective information into a smaller detector format, enabling kilohertz-rate volumetric voltage imaging in the in vivo mouse hippocampus. b, Scanning LFM improves spatial resolution by translating the intermediate image relative to the lenslet array, whereas virtual-scanning LFM replaces mechanical scanning with deep-learning reconstruction, enabling ultrafast voltage imaging in adult Drosophila. c, Confocal LFM combines line illumination with confocal masking to reject out-of-focus background, improving SBR for high-speed in vivo voltage imaging across large neuronal populations in mouse cortex. Adapted with permission from refs.^51,53,54.

Although LFM is scanless, its volumetric rate is often constrained by camera bandwidth: recording full light-field data requires large sensors, which typically operate at lower frame rates and make kilohertz volumetric imaging challenging. To address this “big data” bottleneck, Guo et al. developed a neuromorphic LFM framework that replaces the conventional 2D sensor with an event camera.⁵⁰ Rather than recording full frames, event cameras capture only pixel-wise intensity changes, substantially reducing data volume and better utilizing bandwidth to enable kilohertz rates. A recent work demonstrated that this frame-free approach can resolve both single action potentials and fast network dynamics in cultured neurons (Forli, Angelo, et al., bioRxiv, 2026). However, current event cameras remain limited for voltage imaging by several practical factors, including lower pixel counts than modern scientific CMOS cameras, finite contrast sensitivity thresholds, bias-dependent background activity arising from photon and electron shot noise (Graca, Rui, Brian McReynolds, and Tobi Delbruck, arXiv, 2023), and reduced performance for small-amplitude analogue signals near the detection threshold. As a result, although they are well suited for capturing sparse, fast transients, they are currently less effective for faithfully recording weak continuous voltage fluctuations such as subthreshold oscillations. These limitations are not entirely fundamental: ongoing improvements in pixel density, back-side illumination, fill factor, contrast sensitivity, and readout architecture should further expand their utility for voltage imaging.

A complementary strategy to reduce data burden is compressive light-field acquisition. Wang et al. introduced squeezed light-field microscopy (SLIM),⁵¹ motivated by the observation that perspective images in conventional LFM are highly redundant, differing primarily by depth-dependent disparities across subaperture views.⁵² SLIM exploits this redundancy by compressing perspective measurements along selected orientations and redistributing them across views, greatly reducing dataset size while enabling high-resolution 3D imaging on a low-pixel-count, high-speed camera (Fig. 2a).

Beyond data throughput, LFM faces fundamental challenges in spatial resolution. Because LFM partitions the objective pupil across angular views, it inherently trades diffraction-limited lateral resolution for depth information. Several approaches have been proposed to recover resolution. Scanning light-field microscopy reintroduces scanning by translating the intermediate image across the lenslet array (Fig. 2b, upper panel), leveraging the lenslet-induced diffraction structure to recover high-frequency content.⁴⁸ While effective, the required scanning can introduce motion artifacts in highly dynamic samples. To avoid physical scanning, Lu et al. developed virtual-scanning light field microscopy, which replaces mechanical scanning with a physics-guided deep-learning reconstruction.⁵³ By exploiting phase correlations across angular views, their virtual-scanning network achieved a near diffraction-limited resolution (Fig. 2b, middle and lower panels).

The speed advantage of LFM comes with important trade-offs. LFM is inherently tomographic, acquiring information through nonuniform sampling in spatial-frequency space that typically oversamples low frequencies while undersampling high frequencies. This makes it intrinsically less efficient than direct spatial-domain scanning for uniform, high-fidelity recovery of fine details. Moreover, because tomographic measurements are multiplexed, signals from different locations are mixed on the detector, reducing the effective dynamic range. If bright and dim objects are present in the same FOV, shot noise from the bright object can propagate through the reconstruction and contaminate the dim object, and this effect cannot be removed numerically after detection. More broadly, like other point-spread-function-engineering approaches, LFM encodes 3D position into a 2D camera pattern by spreading photons over many pixels. While reconstruction can computationally reassign these signals to their object locations, the accumulated detector noise is also propagated, ultimately reducing SNR.

Finally, unlike confocal microscopy and LSFM, LFM lacks intrinsic optical sectioning. Although computational refocusing can reconstruct depth planes, out-of-focus fluorescence cannot be physically rejected, which reduces contrast—especially in deeper tissue or when background consumes a large fraction of the camera dynamic range. To address this, Bai et al. adopted confocal light-field microscopy (confocal LFM), which integrates confocal line detection with LFM to suppress background fluorescence (Fig. 2c).^54,55 With 3D deconvolution, confocal LFM achieved an improved SNR by ~2.4× relative to widefield LFM. More recently, Lu et al. introduced line-confocal illumination in sLFM⁵⁶, achieving diffraction-limited, high-speed 3D imaging with SBR performance comparable to spinning-disk confocal microscopy by synchronizing line illumination with the camera rolling shutter.

3. STATE-OF-THE-ART TWO-PHOTON VOLTAGE IMAGING

Compared with 1P excitation microscopy, 2P voltage imaging enables deeper imaging because of reduced optical scattering (Box 2), allowing interrogation of neural dynamics in deep brain regions. However, owing to the much smaller absorption cross-section, 2P voltage imaging faces greater challenges in balancing speed and SNR. In this section, we review state-of-the-art 2P approaches in three categories: fast scanning, spatiotemporal multiplexing, and scanless holography, each with distinct trade-offs among speed, SNR, and excitation power. Representative performance metrics are summarized in Table 1.

3.1. FAST SCANNING

Like 1P confocal microscopy, 2P excitation microscopy typically relies on point scanning to form 2D or 3D images. Conventional raster scanning with galvanometric mirrors, however, is generally limited to only tens of frames per second—far too slow to resolve millisecond-scale membrane voltage dynamics. Faster scanners, including polygon scanners and resonant galvos, can substantially increase throughput,⁵⁷ albeit over a highly restricted stripe-shaped FOV. This fundamental speed–FOV trade-off can be partially alleviated using an optical scan multiplier unit,⁵⁸ which passively rescans a single line using a lenslet array and a retroreflecting mirror to increase effective scan throughput.⁵⁹

Beyond simply increasing scan speed, throughput can also be improved by reducing the number of samples required per frame. One route is to trade spatial resolution for speed by laterally expanding the excitation spot to approximate neuronal soma dimensions, while maintaining axial confinement via temporal focusing.⁶⁰ A more aggressive reduction in sampling abandons continuous rasterization in favour of targeted readout. For instance, acousto-optic deflectors (AODs) enable random-access targeting by repositioning the focal spot within microseconds, allowing kilohertz-rate sampling at preselected locations (Fig. 3a).^61,62

Figure 3. High-speed 2P voltage imaging.

Advanced 2P microscopy uses targeted excitation, compressive sampling, multiplexing and scanless holography to overcome the speed limits of raster scanning. a, Random-access scanning with acousto-optic deflectors (AODs) targets predefined cellular locations, enabling high-temporal-resolution voltage imaging in organotypic rat hippocampal slice cultures. b, Scanned line angular projection microscopy (SLAP) uses tomographic line projections and single-pixel detection to reconstruct sparse neural activity with fewer measurements. c, FACED-based spatiotemporal multiplexing generates temporally staggered beamlets to multiply scan throughput, enabling kilohertz voltage recordings in the visual cortex of awake mice; black traces show individual neuronal signals and the purple trace highlights burst firing. d, Scanless holographic 2P imaging uses computer-generated holography and temporal focusing to illuminate multiple targets simultaneously, improving dwell time and SNR. EOM, electro-optic modulator; M, mirror; NPBS, non-polarizing beam splitter; PBS, polarizing beam splitter; SLM, spatial light modulator Adapted with permission from refs.^61,73,74,78.

These performance gains come with limitations in scalability and robustness. The limited bandwidth of AODs constrains how many targets can be addressed while maintaining the high sampling rates required for voltage imaging, limiting scalability to large neuronal populations. Emerging alternatives, such as one-dimensional phase modulators⁶³, and ongoing developments including AOD-based holography⁶⁴ may help expand effective bandwidth. In addition, because random-access strategies acquire relatively little spatial context, they are more susceptible to motion artifacts in freely behaving animals, where movement cannot be retrospectively corrected via image registration. Potential remedies include 3D ribbon scanning⁶⁵, integration of real-time motion tracking⁶⁶, and generation of multiple spots to create an excitation volume larger than a neuron’s typical movement range².

In parallel, reducing the average optical power required for high-speed 2P voltage imaging is a central objective, both to mitigate heating and to improve photostability. For most currently available voltage indicators, 2P excitation typically requires ~10,000-fold higher average laser power per cell to achieve photon counts comparable to 1P excitation.⁶⁷ One strategy to reduce the required average power is to illuminate only when the scan traverses predefined regions of interest (ROIs)—for example, neuronal somata or membranes—analogous to targeted illumination in widefield microscopy.⁶⁸ In principle, such temporal gating could be implemented by modulating the laser with an external electro-optic modulator (EOM); however, because gating would occur after amplification, maintaining sufficient pulse energy during the “on” windows can demand impractically high laser power. To overcome this limitation, Li et al. developed an adaptive excitation source (AES) that integrates an internal EOM to gate the 2P seed laser prior to amplification, with a feedback loop to stabilize output.⁶⁹ By concentrating excitation only on ROIs during scanning, AES reduces overall sample exposure by more than an order of magnitude (Zhao, Shitong, et al., Research Square, 2024).

Finally, even in the absence of scanner or laser constraints, the maximum pixel rate in raster-scanning 2P microscopy is fundamentally limited by the fluorescence lifetime, typically several nanoseconds, which sets the minimum pixel dwell time needed to avoid temporal crosstalk between adjacent pixels.⁷⁰ This lifetime-imposed dwell-time floor restricts the total pixel budget per frame (or volume) at kilohertz acquisition rates, thereby forcing a trade-off between speed and spatial sampling density and ultimately reducing spatial resolution over large 2D or 3D FOVs.

3.2. SPATIOTEMPORAL MULTIPLEXING

Beyond increasing scan speed, 2P imaging throughput can be boosted by spatiotemporal multiplexing, in which multiple excitation sites are interrogated per unit time. In the spatial domain, one implementation is parallel scanning, analogous to spinning-disk confocal microscopy: an array of excitation foci is rastered across the FOV while fluorescence from multiple spots is detected in parallel using a detector array.⁷¹

A conceptually different form of spatial multiplexing replaces point-wise sampling with projection-based (tomographic) measurements that encode signals across the full FOV. When combined with compressive sensing principles, such approaches can substantially reduce the number of measurements required for accurate reconstruction. Kazemipour et al. demonstrated this concept with scanned line angular projection microscopy (SLAP), in which four angular projections are acquired using only four line scans and the image is reconstructed using a simple filtered back-projection algorithm (Fig. 3b).^72,73

Multiplexing can also be encoded in time, enabling single-detector readout without a detector array. Wu et al. introduced free-space angular-chirp-enhanced delay (FACED), a module comprising a pair of slightly misaligned mirrors that converts a single laser beam into a one-dimensional (1D) array of beamlets (Fig. 3c, upper panel).⁷⁴ The timing offset among these beamlets enables fluorescence from each focus to be read out sequentially using a single photodetector. By sweeping the 1D beamlet array along the orthogonal axis with a galvanometric mirror, the FACED 2P microscope achieved kilohertz full-frame imaging (Fig. 3c, lower panel).⁷⁵ More recently, Liu et al. extended the FACED principle using a diffractive beam splitter to generate a 16-beamlet 1D array (Liu, Zheng, et al., bioRxiv, 2025) significantly expanding the effective FOV.

Finally, multiplexing schemes can combine temporal and spatial encoding with parallel detection. For example, Platisa et al. generated eight excitation beamlets from a single laser source and detected the resulting fluorescence in parallel with a linear photomultiplier tube array.⁷⁶ Nonetheless, these multiplexed architectures share practical limitations, including increased system complexity, reduced flexibility in wavelength, FOV, and speed tuning, and sensitivity to pulse-energy fluctuations.

3.3. SCANLESS HOLOGRAPHY

In contrast to scanning-based strategies, scanless 2P imaging uses computer-generated holography in combination with temporal focusing to simultaneously deliver excitation to multiple predefined sites (e.g., neuronal somata or membranes) across the FOV (Fig. 3d). By eliminating raster scanning, this approach maximizes excitation dwell time at each target and can improve SNR.⁷⁷ Current implementations commonly pair holographic excitation with widefield detection, which offers substantially higher acquisition rates than conventional point-detector readout, albeit with increased vulnerability to tissue scattering.⁷⁸ Notably, scanless 2P approaches appear particularly well matched to emerging VSD- and rhodopsin-based voltage indicators⁷⁹, positioning them as a promising direction for large-scale, high-speed voltage imaging.

4. OPEN CHALLENGES AND OPPORTUNITIES

4.1. DEEP 1P VOLTAGE IMAGING

Deep brain imaging is essential for understanding neuronal activity in areas like the hippocampus, crucial for memory and learning. One approach to increase the imaging depth, whilst maintaining the brain intact, is to utilize longer wavelength excitation or emission light, as these wavelengths experience reduced scattering and absorption in biological tissues. However, voltage imaging in the NIR region (NIR-I: ~700–900 nm; NIR-II: ~1000–1700 nm) still faces substantial challenges, primarily due to the scarcity of effective voltage indicators in these longer wavelength ranges. Compared to well-established visible-range indicators, robust high-performance NIR voltage indicators remain scarce⁸⁰, largely because of fundamental photophysical constraints.⁸¹ Recent efforts, such as the development of Archon1⁸² and nirButterfly⁸³, have shown progress in extending voltage indicators into the NIR-I range, but these indicators are still in the early stages of development and lack the maturity and widespread adoption of their visible-range counterparts.⁸⁴

Another critical challenge is the performance limitation of detectors suitable for imaging at longer wavelengths. For example, current camera sensors designed for NIR-II imaging, such as Indium Gallium Arsenide (InGaAs) cameras, typically suffer from low pixel counts due to high fabrication costs and technological complexity. They also exhibit substantially higher readout and dark-current noise than silicon-based CCD or CMOS detectors, limiting their suitability for detecting weak voltage signals.⁸⁵

Currently, the majority of 1P voltage imaging with single cell resolution in vivo, including confocal strategies, rely on the detection of ballistic or quasi-ballistic photons from sparsely labelled neuronal populations. With dense labelling, scattered fluorescence creates broad, overlapping spatial footprints, making it difficult to assign activity to individual cells. Spatiotemporal demixing methods could help overcome this limitation. For example, activity localization imaging uses the relative sparsity of action potentials to identify neuronal footprints beyond the ballistic scattering limit.⁸⁶ Other approaches exploit speckle contrast to separate temporally varying signals in diffusive scattering regimes.^87,88 Although demonstrated mainly in 1P calcium imaging⁸⁹, these strategies could help extend the imaging depth of camera-based 1P and 2P voltage imaging.

4.2. PHYSIOLOGICALLY TOLERABLE 2P VOLTAGE IMAGING

As outlined in Section 3, improving 2P voltage imaging requires advances in imaging depth, accessible volume, spatiotemporal resolution, and the number of neurons recorded simultaneously. The main constraint is the maximum excitation power that can be delivered to brain tissue without excessive heating.^90,91 Because 2P excitation requires substantially higher power per cell than one-photon excitation,⁶⁷ and even higher power at depth to compensate for scattering, current methods face a stringent trade-off between shot-noise-limited sensitivity and photothermal damage. Reducing the average power needed to achieve sufficient SNR is therefore a central goal.

Several strategies could help. Next-generation laser sources with high pulse energy, short pulse durations (<70 fs), tuneable repetition rates and wavelengths matched to voltage indicators could improve excitation efficiency while reducing heating. Restricting excitation to membrane-localized indicators could further lower power requirements, although this would require robust real-time motion correction in vivo. Polarization control offers another underexplored route, as aligning excitation polarization with fluorophore transition dipoles can enhance 2P excitation efficiency.^92,93 Finally, improved voltage indicators with larger fractional fluorescence changes would directly increase signal per photon. Together, these advances could make 2P voltage imaging more practical for recording biologically relevant neuronal ensembles in deep brain regions.

4.3. SCALING VOLTAGE IMAGING TO MESOSCALE FIELDS OF VIEW

Beyond depth, a major frontier is scaling voltage imaging laterally from local circuits to mesoscale networks. Because brain function emerges from distributed communication across distant regions⁹⁴, large-FOV voltage imaging could reveal fast activity propagation, inter-regional coupling and coordinated population dynamics, bridging cellular electrophysiology with systems-level behaviour and cognition.

Despite recent advances, most voltage imaging demonstrations remain limited to fields of view of hundreds of micrometres, whereas calcium imaging routinely reaches millimetre-scale coverage across multiple cortical regions.^95–97 This gap leaves the systems-level potential of cellular-resolution voltage imaging largely underexplored.

For camera-based approaches, mesoscale imaging is mainly constrained by detector bandwidth: kilohertz imaging over larger sensor areas requires substantially higher pixel throughput. Encouragingly, camera performance has improved rapidly and is likely to continue doing so. In parallel, architectural strategies such as camera arrays^96,98 and computational strategies such as compressive sensing^51,73,99 offer additional routes to increase effective throughput. A second constraint is the photon budget: high-speed voltage imaging is typically shot-noise limited, making photon collection efficiency paramount and motivating the use of high-NA mesoscope objectives. Recent efforts such as the Cousa objective¹⁰⁰ and the RUSH platform⁹⁶ extend FOVs but were primarily optimized for calcium imaging and typically operate at NA < 0.5. For voltage imaging, where preserving photons is critical, there is a need for objective designs that simultaneously deliver high NA, expanded FOV, and compatibility with in vivo preparations.

For scanning-based approaches, particularly 2P microscopy, mesoscale voltage imaging is limited by scanner throughput. Faster scanning often requires smaller apertures or reduced scan angles, compromising resolution and FOV.^58,101 Although fast scanners and multiplexed strategies continue to improve, no current method yet provides kilohertz-rate 2P voltage imaging at true mesoscale coverage.

4.4. FROM HEAD-FIXED TO FREELY MOVING: MINIATURIZATION

Finally, realizing network-scale voltage imaging during naturalistic behaviour will require not only larger FOVs on the benchtop, but also the ability to shrink these capabilities into head-mounted form factors. Most current voltage imaging systems rely on advanced, highly sensitive, and high-speed camera sensors, as well as powerful excitation lasers. Such systems can accommodate studies involving freely behaving mice only in head-restrained configurations, wherein the animal’s head is fixed relative to the imaging objective while still allowing natural limb and body movements. In contrast, calcium imaging has successfully transitioned from conventional benchtop systems to miniaturized, head-mounted microscopes, allowing mice complete freedom of movement during imaging.^102,103 This significant advancement has greatly expanded the scope of behavioural neuroscience studies, enabling detailed investigations of neural activity during complex behaviours, social interactions, and navigation tasks.

For widefield or targeted-excitation voltage imaging, miniaturization is particularly challenging because it requires camera sensors that combine high frame rate with high sensitivity and low noise. Scientific-grade CMOS/CCD detectors typically remain too large and heavy for direct head mounting. Recently, MiniVolt integrated state-of-the-art CMOS sensors into miniature platforms, and demonstrated in vivo voltage spike recordings,¹⁰⁴ but the resulting device is still relatively heavy for freely moving mice (16.4 g). One alternative is to use fibre-optic image bundles to relay fluorescence from the head to a remote detector, enabling the bulky camera to remain off-animal.^105,106 This architecture preserves only essential optics on the head, substantially reducing payload and facilitating multimodal integration without increasing head-borne weight.¹⁰⁶ Although commercial image fibre bundles typically contain only a few thousand cores—limiting resolution for fine neurites—they can still support soma-level recordings. Extending such relay architectures to voltage imaging, for example by pairing image fibre bundles with high-sensitivity cameras, could enable multi-target voltage recordings in freely moving animals.

Scanning-based architectures face a different set of miniaturization constraints. Scanning miniscopes commonly use fibre relays to place the laser source and detector off-animal, leaving only compact imaging optics and a miniature scanner, such as a micro-electro-mechanical systems (MEMS) mirror, on the head.^107,108 This layout is, in principle, compatible with voltage imaging. However, as in tabletop systems, imaging speed and FOV are ultimately limited by scanner throughput. Adapting strategies such as spatiotemporal multiplexing¹⁰⁹ and multi-region imaging¹¹⁰ could help increase throughput in future miniaturized voltage-imaging platforms.

5. INTEGRATIVE PERSPECTIVES ON THE TECHNOLOGICAL LANDSCAPE

The recent progress reviewed above makes clear that no single microscopy architecture is universally optimal for voltage imaging. Instead, the best-performing systems emerge from co-design across three tightly coupled layers: the voltage indicator, the imaging instrument, and the downstream inference pipeline. Voltage indicators determine the achievable signal amplitude, spectral window, membrane localization, photostability, and kinetics; the microscope determines photon throughput, background rejection, spatial coverage, and sampling speed; and the analysis pipeline determines how efficiently weak optical signals can be converted into biologically meaningful readouts such as spikes, subthreshold dynamics, propagation patterns, or functional connectivity. Framing the field through this joint design lens helps clarify why different imaging modalities excel in different regimes and where the next technical opportunities lie.

5.1. TASK-DEFINED PERFORMANCE METRICS: SNR, SBR, AND SPATIAL SAMPLING

In voltage imaging, performance should be judged relative to the downstream biological question rather than by frame rate alone. For many experiments, the relevant objective is not simply to maximize raw photon counts, but to maximize the fidelity of a specific inference task – for example, detecting single spikes, resolving subthreshold oscillations, separating neighbouring neurons, or tracking activity across depth. Under the shot-noise-limited conditions that characterize many voltage imaging experiments, SNR is fundamentally constrained by photon collection efficiency, making optical throughput a primary design consideration. In this respect, parallel or snapshot detection schemes offer a major advantage because they collect photons from many spatial locations simultaneously rather than sequentially. This “snapshot advantage” is especially valuable when signals are fast, dim, and distributed across a broad field.^111,112

By contrast, SBR is primarily determined by background rejection. Here, optical-sectioning methods such as confocal microscopy, light-sheet microscopy, and 2P excitation are fundamentally favoured because they physically suppress out-of-focus fluorescence before it contributes shot noise to the measurement. Widefield and conventional light-field approaches can partially mitigate background through sparse labelling, targeted illumination, deconvolution, or computational unmixing, but they do not remove the photon noise generated by out-of-focus light that has already reached the detector. This distinction is important: numerical reconstruction can improve contrast in the reconstructed image, but it cannot recover SNR lost to background shot noise at acquisition. Accordingly, one of the most promising design directions is to combine widefield-style parallel detection with physical background suppression, as exemplified by confocal light-field implementations and related hybrid architectures.

Spatial resolution must likewise be interpreted in the context of the intended measurement. For single-cell physiology, the goal may be to resolve soma, dendrites, axons, or conduction dynamics along fine neuronal processes, which pushes microscopy toward high NA, strong optical sectioning, and often volumetric capability. For population imaging, by contrast, the dominant requirement is often reliable demixing of nearby neurons over a larger field and longer duration, even if the system does not preserve submicron detail everywhere. These two regimes place different demands on microscope design and should not be treated as interchangeable use cases.

5.2. INDICATOR-INSTRUMENT CO-DESIGN

The properties of voltage indicators should play a more central role in determining imaging architecture. Indicators differ widely in brightness, membrane trafficking, dynamic range, kinetics, photostability, excitation/emission spectra, and susceptibility to background from unresponsive or mislocalized molecules. These properties directly affect whether a modality is limited primarily by photon budget, optical crosstalk, depth penetration, or acquisition bandwidth. For example, dimmer but fast indicators place a premium on optical throughput and may favour parallel detection, whereas indicators with strong membrane confinement and high brightness can better exploit targeted or multiplexed strategies. Likewise, red-shifted or near-infrared indicators could expand opportunities for deeper 1P imaging or reduced scattering, but only if detector sensitivity, optical coatings, and excitation sources are co-optimized accordingly. Thus, microscope design should not be viewed as independent of indicator chemistry; rather, indicator development and instrument development should proceed as a coupled optimization problem.

This co-design principle is especially important in 2P voltage imaging. Because 2P excitation is intrinsically photon inefficient compared with 1P excitation, successful systems often rely on strategies that reduce the number of sampled spatial points, increase dwell time at relevant locations, or multiplex excitation in space and time. These choices are only worthwhile if they remain compatible with the photophysics and localization properties of the chosen indicator. Similarly, in 1P imaging, targeted illumination is most beneficial when labelling density, membrane specificity, and structural priors make it possible to illuminate only informative regions. In both cases, the practical question is not simply which microscope is fastest, but which microscope is best matched to the signal model defined by the indicator and the biological sample.

5.3. CAMERA TECHNOLOGY, DATA BURDEN, AND CO-DESIGN WITH COMPUTATION

For camera-based voltage imaging, detector bandwidth has historically been a major bottleneck, particularly for large-FOV or volumetric measurements. That constraint is now shifting. Modern scientific cameras increasingly offer combinations of high frame rate, large pixel count, and high sensitivity that make kilohertz-class recording feasible for selected imaging geometries. As a result, the more consequential bottleneck in many settings is no longer raw sensor performance alone, but the generation, transfer, storage, and analysis of massive time-resolved datasets. This challenge becomes especially severe for long-duration recordings and for 3D methods, where data scale rapidly in both time and depth.

This shift creates an opportunity for deeper co-design between imaging hardware and computational inference. Rather than treating image formation and analysis as separate stages, future systems can be designed so that the acquisition strategy directly reflects the intended inference task. Compressive measurements, optical segmentation, sparse ROI-matched readout, event-driven sensing, and temporally coded acquisition all represent versions of the same idea: do not measure redundant information if the downstream task does not require it. Voltage imaging is particularly amenable to this philosophy because neural activity is often sparse in space, sparse in time, or structured by anatomical priors. Methods such as projection-based imaging, compressive light-field acquisition, and segmentation-guided recording illustrate how reducing measurement dimensionality at the hardware level can substantially ease the computational and storage burden while preserving biologically relevant information. In this sense, the field may benefit less from ever larger raw datasets and more from task-aware acquisition coupled to model-based or learning-based reconstruction.

6. CONCLUSION REMARKS

High-speed fluorescence microscopy, coupled with recent advancements in voltage indicators, has significantly expanded the capacity to study fast neuronal dynamics with high spatial and temporal resolution. While calcium imaging remains a valuable tool, voltage imaging offers a direct and precise measurement of membrane potential fluctuations, enabling a deeper understanding of neural circuit computations.

A missed opportunity in the field is the tendency to frame voltage imaging primarily as a technology for large-scale population recording. While that direction is clearly important, voltage imaging also has tremendous potential for single-cell and subcellular physiology, where the design priorities may be quite different. Imaging back-propagating action potentials, dendritic integration, axonal conduction, or compartment-specific synaptic responses demands high spatial resolution, high SBR, and often 3D coverage over elongated structures. These applications may favour confocal, light-sheet, or 2P approaches, potentially combined with extended-depth-of-field strategies, Bessel beams, or other optical designs that better accommodate the intrinsically 3D geometry of neuronal processes. Closer consideration of these distinct biological needs can help users map specific neuroscience questions onto the most appropriate instrument classes, rather than treating all voltage imaging modalities as competitors in a single performance race.

Similarly, the technological landscape could be broadened to include non-imaging readout formats, such as fibre-based or photometry-like voltage measurements. Although these approaches sacrifice spatial information, they can offer substantial advantages in simplicity, stability, depth accessibility, and compatibility with freely behaving preparations. They may be particularly useful when the scientific question centres on population-average voltage dynamics in a genetically defined cell class rather than on resolving individual neurons. Integrating such modalities into the broader voltage-imaging conversation would help unify currently fragmented application domains and emphasize that the central objective is not imaging per se, but robust optical access to fast membrane potential dynamics across scales and preparations.

Overall, the next phase of the field will likely be defined less by isolated improvements in frame rate, resolution, or FOV and more by system-level co-optimization: matching indicator properties to optical architectures, matching acquisition strategies to inference tasks, and matching instrument specifications to the biological scale of interest. By organizing the field around shared constraints and design principles, rather than only around modality labels, we can better identify which trade-offs are fundamental, which are engineering contingencies, and where the most impactful opportunities for innovation now lie.