1.

Introduction

Cell migration and single cell metabolism are popular areas of investigation across immunology, cancer research, and developmental biology.¹ However, a significant bottleneck exists in 3D imaging of metabolic features in moving cells throughout a live model organism, such as the zebrafish. Current methods require sample destruction (e.g., mass spectrometry, flow cytometry, and histology) that removes 3D context and prevents time-course studies that follow the fate of the same cell, or fluorescent reporters that require sample manipulation.² Additionally, fluorescent reporters often provide a binary on/off indicator of expression rather than a continuous variable of dynamic cell state, even though most cells are known to function on a spectrum of activity.

The reduced form of nicotinamide adenine dinucleotide (phosphate) [NAD(P)H] is a naturally fluorescent metabolic co-factor involved in hundreds of reactions within the cell (NADH and NADPH are optically indistinguishable and their fluorescence is collectively referred to as NAD(P)H).^2–13 The fluorescence lifetime of NAD(P)H is distinct in the free (short lifetime) and protein-bound (long lifetime) conformations, so fluorescence lifetime imaging microscopy (FLIM) of NAD(P)H provides information on protein-binding activities, preferred protein-binding partners, and other environmental factors, such as pH and oxygen, at a single cell level.^{2,10,12,14,15} FLIM of NAD(P)H is advantageous for imaging samples that are difficult to label with fluorescent reporters, for in vivo imaging of preclinical models, and for visualizing fast biological processes. For example, NAD(P)H FLIM can probe the temporal and spatial regulation of macrophage metabolism during tissue damage and repair in live zebrafish.¹¹ NAD(P)H FLIM typically relies on two-photon (2P) laser scanning microscopy, which provides intrinsic optical sectioning, improved penetration depth, and high spatial resolution with minimal phototoxicity.^16,17 However, autofluorescence FLIM traditionally suffers from long acquisition times and limited spatial views, partly due to the low fluorescence quantum yield (${\mathrm{\Phi }}_f$ for NAD(P)H is between 0.02 and 0.10 depending on protein binding) as well as lower molar absorption coefficient [$ϵ$ for NAD(P)H is $6200{\mathrm{cm}}^{-1}{\mathrm{M}}^{-1}$], which makes the autofluorescence signal hundreds of times dimmer than that of conventional fluorophores (e.g., the green fluorescent protein [GFP] has ${\mathrm{\Phi }}_f=0.79$ and $ϵ=25500{\mathrm{cm}}^{-1}{\mathrm{M}}^{-1}$).^18–20 There is a significant opportunity to develop new NAD(P)H FLIM imaging hardware to better understand 3D cellular dynamics in vitro and in vivo.

Single-photon avalanche diodes (SPAD) are solid-state photodetectors with high photon counting and time-resolving capabilities.²¹ SPADs are semiconductor p–n junctions reverse-biased above their breakdown voltage such that a single-photon event can create an electron–hole pair and trigger a self-sustaining avalanche current. Advances in SPAD technology have led to the development of multipixel arrays with on-chip integrated timing electronics [such as time-to-digital converters (TDC)] capable of distinguishing single-photon arrival times.²¹ SPAD arrays have numerous applications in biophotonics, including fluorescence correlation spectroscopy,^22–24 Raman spectroscopy,^25–27 spectrally resolved FLIM,²⁸ and optical tomography.^29–31 Larger SPAD arrays ($>32\times 32$) have also been used extensively for FLIM in combination with multifocal multiphoton excitation, providing exceptional speed advantages while maintaining the spatial and temporal resolution of traditional single-beam scanning microscopes.^32–35 Other application examples include real-time widefield FLIM with time-correlated single photon counting (TCSPC) or time-gated sensing to image standard fluorescence solutions and quantum dots,^36,37 fluorescently labeled cells,^38,39 and fungal spores.⁴⁰ High-speed time domain FLIM with SPAD arrays has been used for rapid tumor phantom margin measurements⁴¹ and to distinguish labeled vasculature⁴² and melanomas in live mice.⁴³ Other geometries and biophotonics applications are detailed in the extensive review by Bruschini et al.²¹

Although SPAD arrays can be easily integrated into widefield microscopes, these microscopes lack optical sectioning capabilities and therefore offer poor sensitivity for NAD(P)H FLIM, which targets a weak autofluorescence signal over a high background. Light-sheet microscopy, also known as selective plane illumination microscopy (SPIM), uses a light sheet perpendicular to the imaging axis to illuminate the focal plane and achieve optical sectioning.⁴⁴ Light-sheet microscopy is an attractive alternative to widefield or laser scanning systems because it provides 3D optical sectioning along with fast volumetric imaging due to the following features: first, SPIM enables higher throughput due to parallel acquisition of image pixels; second, out-of-focus sample exposure and photobleaching are eliminated; finally, SPIM provides better excitation efficiency and lower irradiance compared to scanning a diffraction-limited spot in epi-illumination geometry, in turn lowering the light dose and phototoxicity relative to single-photon confocal systems.^44–46 Light-sheet FLIM has been performed on GFP-labeled cancer cell spheroids and C. elegans using microchannel plate photomultiplier tube detectors⁴⁷ and live transgenic zebrafish using a frequency-domain two-tap CMOS FLIM camera.⁴⁸ Gated optical image intensifiers have been used in conjunction with CMOS⁴⁹ or CCD⁵⁰ cameras on SPIM-FLIM systems to image GFP-labeled live zebrafish⁴⁹ or canine kidney cysts,⁵⁰ respectively. To our knowledge, autofluorescent NAD(P)H FLIM has not yet been performed using light-sheet systems.

Here we demonstrate the first use of a commercially available SPAD array for detecting NAD(P)H autofluorescence in a light-sheet geometry. To improve image quality, the SPAD image from the FLIM arm of the light-sheet system was upscaled with the CMOS image from the intensity arm of the light-sheet system. The sensitivity of the NAD(P)H light-sheet FLIM system was confirmed with metabolic perturbations to pancreatic cancer cells with 10 s integration times, and the NAD(P)H light-sheet FLIM system was demonstrated in vivo with live neutrophil imaging in a wounded zebrafish tail, also at 10 s integration times. Finally, the theoretical and practical imaging speeds for NAD(P)H FLIM are compared across laser scanning and light-sheet geometries, indicating a $30\times$ to $6\times$ acquisition speed advantage for the light-sheet compared to the laser scanning geometry.

2.

Methods

2.2.

Multidirectional Selective Plane Illumination Microscopy with SPAD Array

The mSPIM system is detailed in Fig. 1,⁵² with multidirectional capabilities maintained in the intensity arm only. Illumination was achieved using two opposing light sheets. Three $10\times /0.3\mathrm{NA}$ water immersion objective lenses (CFI60, Nikon Instruments, Inc.) were used for illumination and detection. The “Flamingo” T-SPIM design⁵³ was modified to incorporate a QuixX 375-70PS picosecond-pulsed diode laser (Omicron-Laserage Laserprodukte GmbH) in the right illumination arm for NAD(P)H excitation. The laser produced 90 ps pulses at a 50 MHz repetition rate with an average power of 0.4 mW at the sample. The output beam had a diameter of 1 mm which was used along with a $f=100\mathrm{mm}$ cylindrical lens (Thorlabs) to underfill a $10\times /0.3\mathrm{NA}$ objective lens, generating a $200\mu \mathrm{m}$-wide light sheet with $9\mu \mathrm{m}$ waist thickness that matched the FLIMera sensor size (right arm, Fig. 1). A bank of continuous-wave lasers (TOPTICA) was used to generate a 1.5 mm-wide light sheet (left arm, Fig. 1) for targeting other fluorescent labels for intensity imaging using a larger sCMOS sensor ($13.3\mathrm{mm}\times 13.3\mathrm{mm}$) with $2048\times 2048\text{pixels}$, $6.5\mu \mathrm{m}$ pitch (Panda 4.2, PCO GmbH). Brightfield trans-illumination was achieved using a red LED and imaged using the sCMOS sensor.

Fig. 1

Schematic of the light-sheet FLIM microscope (adapted with permission from Ref. 52). (a) Top view of the SPIM system. The right illumination arm uses a 375 nm ps-pulsed diode laser operated at 50 MHz repetition rate to excite NAD(P)H fluorescence via a $200\mu \mathrm{m}$ wide light sheet, while the time-resolved SPAD camera images the NAD(P)H autofluorescence lifetime (after a 495 nm LP DM and a $440/80\mathrm{nm}$ filter) and the sCMOS camera images the NAD(P)H intensity (using a $530/55\mathrm{nm}$ filter). The left illumination arm uses a bank of CW lasers to excite other fluorophores via a 1.5 mm wide light sheet, while emissions are imaged on the sCMOS camera. A red LED provides trans-illumination for bright field imaging. (b) Side view of the beams in the illumination arms and the orientation of the sample tube. DM, dichroic mirror.

NAD(P)H was excited at 375 nm (pulsed laser, 50 MHz, integration time = 10 s on SPAD array) and mCherry at 561 nm (CW laser, integration = 1 s on sCMOS camera). Emissions were split by a 495 nm long-pass dichroic mirror (DM); NAD(P)H was captured using a 440/80 nm bandpass filter on the SPAD array; and a lower NAD(P)H intensity signal was collected through a $530/55\mathrm{nm}$ bandpass filter on the sCMOS camera (integration time = 1 s). Fluorescence from mCherry and brightfield images was captured through a $650/60\mathrm{nm}$ bandpass filter on the sCMOS camera using illumination from the 561 nm CW laser light sheet or the red LED brightfield, respectively.

3.

Results

3.1.

System Characterization and Calibration

The combined IRF of the 50 MHz pulsed laser source and the SPAD array camera is shown in Fig. 2. The system IRF has an FWHM of 380 ps. Due to non-zero DCRs of the SPAD pixels, the measured IRF drops to a non-zero constant in later time bins. This tail of the measured IRF was used to estimate per-pixel DCRs. A histogram of the estimated DCRs for this SPAD camera is shown in Fig. S1 in the Supplementary Material. The median DCR of the SPAD pixels was 34 cps.

Fig. 2

IRF. To measure the combined IRF of the excitation laser (Omicron QuixX 375-70PS) and the SPAD camera (Horiba FLIMera), a fraction of the excitation laser pulses (50 MHz pulse repetition rate) was directly imaged onto the FLIMera camera by imaging a retroreflector target in an epi-illuminated widefield microscope setup. The 20-ns time axis consists of 486 time bins of 41.1 ps width. The IRF has an FWHM of 380 ps. The DCR of individual pixels can be measured from the non-zero tail of the pixels’ IRF.

A map of the inter-pixel delays across the SPAD array sensor was estimated from the coumarin 6 fluorescence lifetime image as shown in Fig. 3 using a pixel-wise cross-correlation maximization scheme. If left uncorrected, this non-uniform delay results in distortion artifacts when aggregating decays from different pixels on the sensor (e.g., for pixel binning during lifetime fit analysis). Figure 4 shows the effect of correcting for these inter-pixel delays on the aggregate decay histogram of coumarin 6. The estimated fluorescence lifetime after such correction (2.51 ns) agrees with reported values in the literature.^62–64 Figures S2 and S3 in the Supplementary Material show lifetime images and aggregate decays for YG and BB fluorescent microspheres. The estimated lifetime of 2.15 ns for YG beads and 2.6 ns for BB beads also agree with reported values in the literature.^65,66

Fig. 3

FLIMera SPAD array camera inter-pixel delay map. To compensate for the inter-pixel timing skew and synchronize the rise time of all $192\times 128\text{pixels}$, the normalized cross-correlation of the measured decay for each pixel with a reference decay was calculated. (a) These cross-correlation curves for a representative group of SPAD pixels (each curve is one pixel). The temporal lag value where the normalized cross-correlation is maximized was recorded and mapped as seen in (b) for every pixel of the SPAD array. The “screamer” dead pixels are dark blue and comprise 15% of all pixels. The inter-pixel lag between pixels on the left and right side of the sensor can be as high as 60 time bins (with 41.1 ps time bin width) or $\sim 2.5\mathrm{ns}$, which is substantial. To synchronize the rise time of all pixels, it suffices to apply an opposite circular shift to the decay histogram of each SPAD pixel by the amount given by the delay map above. (c) The decays from two representative SPAD pixels that have different timing lags. SPAD pixel 1 has a lag of 40 time bins or 1.65 ns with respect to the reference decay, whereas SPAD pixel 2 has a lag of zero. To correct the timing skew, the blue curve should be shifted to the left by 40 time bins.

Fig. 4

Fluorescence lifetime fitting of coumarin 6 in ethanol. A saturated solution of coumarin 6 in ethanol was loaded into an FEP sample tube and imaged on the light-sheet SPAD system with 0.4 mW excitation laser power and 3 s integration time using a green emission filter ($550/100\mathrm{nm}$). (a) The blue curve shows the aggregated decay histogram from all array pixels before correcting for the inter-pixel delays. The red curve shows the aggregated decay histogram after such correction. The corrected decay histogram has a faster rise and higher peak, whereas the uncorrected decay histogram has a slower rise and is temporally broadened. (b) The aggregated decay histogram of all pixels after timing skew correction was fit to a single-exponential decay model using an iterative reconvolution algorithm that performs a least-squares minimization of the residuals. A fluorescence lifetime of 2.51 ns is estimated which agrees with reported published values.^62–64

3.2.

Effects of Cyanide on NAD(P)H Metabolism In Vitro

After the addition of 1 mM cyanide, NAD(P)H FLIM images of PANC-1 cells were taken every 3 min (Fig. 5). We observed the expected decrease in NAD(P)H mean fluorescence lifetime [${\tau }_m$, Figs. 5(a)–5(d) and 5(i)], an increase in the free fraction of NAD(P)H [${\alpha }_1$, Figs. 5(e)–5(h) and 5(j)], and an increase in the NAD(P)H fluorescence intensity [Fig. 5(k)], starting at 3 min post-exposure and continuing over time. These changes in NAD(P)H intensity and lifetime parameters were expected,^4,56,58,67 and the range of NAD(P)H ${\tau }_m$ and ${\alpha }_1$ also fall within published values for cells in culture.^57,68 These experiments confirm the sensitivity of this system to NAD(P)H lifetimes and to physiologically relevant changes in NAD(P)H lifetimes. These studies also demonstrate that this NAD(P)H light-sheet FLIM system can perform time-course imaging of single-cell autofluorescence in a fluorescent background of culture medium. Further, we found that 10 s integration time was sufficient to capture enough NAD(P)H photons for accurate lifetime fitting analysis.

Fig. 5

Cyanide treatment of PANC-1 cells confirms the sensitivity of the light-sheet FLIM system to NAD(P)H fluorescence lifetime changes (Video 1, MPEG, 0.5 MB [URL: https://doi.org/10.1117/1.JBO.28.6.066502.s1]). (a)–(d) NAD(P)H mean fluorescence lifetime (${\tau }_m$) images of PANC-1 cells at different time points after treatment with 1 mM sodium cyanide. The sCMOS NAD(P)H intensity image is used to upscale the SPAD array NAD(P)H lifetime image (e)–(h). Corresponding free fraction of NAD(P)H (${\alpha }_1$) at the same time points. (i) Boxplot of NAD(P)H ${\tau }_m$ of image pixels over time shows a rapid drop in mean lifetime within a few minutes of cyanide treatment. (j) NAD(P)H ${\alpha }_1$ of image pixels increases over time. (k) The pixel intensity of NAD(P)H fluorescence, measured by integrating the decay curve at each pixel in the SPAD camera, increases over time with cyanide exposure. White dot shows the median; red horizontal line shows the mean; box encompasses 25th to 75th percentile range; whiskers extend from the box to 1.5 times the interquartile range. Changes in NAD(P)H mean lifetime ($F$-statistic = 9.78, $p=0.026$), free fraction ($F$-statistic = 21.8, $p=0.006$), and mean intensity ($F$-statistic = 113, $p=0.0001$) with time are significant according to the linear trend test.

3.3.

Light-Sheet FLIM of NAD(P)H In Vivo in Larval Zebrafish

We next sought to measure NAD(P)H lifetime in a dynamic, in vivo environment (Fig. 6). NAD(P)H is present in all cell types of a living organism, so we used a zebrafish caudal fin wound model to track mCherry-labeled neutrophils,⁶⁹ innate immune cells recruited to the injury site. To demonstrate light-sheet NAD(P)H FLIM in these dynamic cells, an mCherry intensity image was acquired using light-sheet excitation from the left (CW) illumination arm whose sectioning plane was aligned to perfectly match the right (pulsed) illumination arm that excites NAD(P)H for the FLIM image. NAD(P)H FLIM was simultaneously acquired in a $z$-stack to track the same neutrophils over time. The mCherry fluorescence was used to identify the cells of interest and create masks to isolate NAD(P)H specifically from the neutrophils, as autofluorescence signal is generated from the whole tail fin tissue in the FOV. These studies demonstrate that the NAD(P)H light-sheet FLIM system is sensitive to single cells in vivo, and that combined intensity and FLIM images can be used to isolate quickly moving cells in vivo.

Fig. 6

Light-sheet imaging of mCherry intensity and NAD(P)H lifetimes in live neutrophils in vivo in a zebrafish caudal fin wound model. (a) The intensity channel of the light-sheet FLIM system shows the superimposed brightfield image of 3 dpf wounded zebrafish tail and light-sheet-excited fluorescence intensity of mCherry+ neutrophils at the wound site (Video 2, MPEG, 2.1 MB [URL: https://doi.org/10.1117/1.JBO.28.6.066502.s2]). The white rectangle shows the corresponding field of view of the smaller SPAD array sensor. (b) NAD(P)H mean fluorescence lifetime (${\tau }_m$) image of the tail (from the SPAD array with 10 s integration). The transection wound (yellow dashed line) and the tip of the notochord (red dashed box) are visible in the lifetime image. (c) Corresponding field of view from the cropped sCMOS image shows mCherry+ neutrophils recruited to the wound site. (d) Pixel mask of neutrophil cells from the mCherry channel applied to the NAD(P)H lifetime image. (e) Masked NAD(P)H lifetime map of the neutrophils.

4.

Discussion and Conclusion

Detector arrays in a widefield geometry provide faster FLIM acquisition speed (i.e., shorter integration time per frame) compared to a single-detector multiphoton or confocal laser scanning geometry. Recent developments in SPAD array sensor technology enable widefield FLIM with high TCSPC temporal resolution. These new SPAD cameras can be combined with low-cost pulsed diode lasers for rapid FLIM. Table 1 shows the theoretical advantage in integration time for the SPAD array used in this work over a standard laser scanning FLIM microscope using a single TCSPC module for acquiring an image of the same size. However, the simplest implementation of widefield FLIM, i.e., an epi-fluorescence geometry, can only be applied to bright fluorescently labeled samples due to lack of optical sectioning. In label-free applications, out-of-focus background fluorescence can overwhelm the weak autofluorescence signals from single cells. Thus widefield autofluorescence FLIM has remained impractical.

Table 1

Comparison of theoretical FLIM imaging speed limit between laser scanning and widefield geometries. The saturation count rate for the SPAD camera with per-pixel TDC modules working in parallel is 30× higher than the saturation count rate of a single TCSPC module in laser scanning microscopes. The speed advantage scales with the number of array pixels.

In this work, we combined multichannel detection using a SPAD array camera with light-sheet excitation to minimize the effects of out-of-focus background fluorescence. In addition to providing optical sectioning, light-sheet illumination is more efficient at exciting fluorophores than epi-illumination because photons traveling in the focus plane have a higher probability of encountering a fluorophore. The peak irradiance of a light sheet is also many orders of magnitude lower than a diffraction-limited spot in laser scanning (e.g., confocal) microscopy, which decreases phototoxicity. Artifacts are noticeable in areas of the presented lifetime images where dead “screamer” SPAD pixels (that comprise 15% of all pixels) cluster together. We also acknowledge that a secondary streaking artifact is visible in Figs. 5 and 6 due to shadowing in the light-sheet excitation path. This is caused by absorption in the sample and can be mitigated by pivoting the light sheet (also known as dithering) such that these shadows are averaged out during the integration period.⁵² Such artifacts could be eliminated in future designs using UV-optimized multidirectional/pivoting optics.

The low quantum yield of endogenous fluorophores [e.g., NAD(P)H, FAD]¹⁸ limits the acquisition speed of autofluorescence FLIM because photodamage to cells occurs before the detector count rate saturates. As such, it is helpful to compare the light-sheet SPAD system to a laser scanning system for the imaging settings used and count rates observed for NAD(P)H FLIM in practice. This comparison is made in Table 2 for the specific 2P laser scanning microscope,⁶⁷ the light-sheet SPAD system, and the PANC-1 cells used in this study. The pixel count rate and photons per pixel are reported before any pixel binning. Figure S4 in the Supplementary Material shows a representative two-photon FLIM image of PANC-1 cells. For the same field of view, number of image pixels, and targeted photon count per pixel, the light-sheet SPAD system in Fig. 1 requires $6\times$ less integration time to acquire the NAD(P)H image compared to a 2P laser scanning microscope. Frame rates of SPAD arrays relative to single-pixel laser scanning systems scale with the number of array pixels, when acquiring the same number of image photons. SPAD arrays are based on the same CMOS fabrication process as existing cameras, so the commercial infrastructure for inexpensive and large-scale manufacturing of high-resolution arrays exists.⁷⁰ This has resulted in a rapid improvement in the capabilities of SPAD systems with improved photon detection probabilities and fill factors (e.g., using microlens arrays over SPAD pixels^71,72), which can further reduce integration time per frame in FLIM or reduce excitation light dose for reduced phototoxicity over long imaging sessions, which is an important consideration given the higher phototoxicity of shorter wavelengths.⁴⁵

Table 2

Comparison of practical FLIM imaging speed limit between a standard 2P laser scanning system and the light-sheet system in Fig. 1. For the same field of view size, number of image pixels, and targeted number of emission photons per image pixel, the light-sheet system required 6× less integration time per image.

Existing light-sheet FLIM microscopes perform well for samples with bright fluorophores, but access to weaker autofluorescence lifetimes is challenging due to limitations in sensitivity and temporal resolution of current detector arrays. Here, we show that it is feasible to use a commercial SPAD array in a light-sheet microscope for NAD(P)H FLIM of single cells in vitro and in vivo. The integration time per image for this configuration is 6× faster than that of traditional laser scanning NAD(P)H FLIM microscopes. Faster acquisition speeds will be achieved as SPAD arrays continue to improve in detection efficiency and array size.

The FLIM frame interval in these studies was limited by write-to-disk speed when saving data in raw photon stream mode, which imposed delays between successive frames. Specifically, the FLIMera camera can only transfer data in photon streaming mode over USB 3.0, with histogram compilation and lifetime analysis on the computer CPU. The process of transferring individual photon time tags costs both USB data transfer bandwidth and analysis time, which prohibits the acquisition of successive frames at the maximum speed of the hardware. Some SPAD array developers use on-chip histogramming with an FPGA⁷³ to substantially reduce this data transfer and analysis bottleneck, enabling larger SPAD arrays with per-pixel TDC. Alternatively, a time-gating scheme with a user-selectable number and width of time gates can be used instead of TCSPC. This approach removes per-pixel TDC electronics and simplifies the design to provide larger SPAD arrays with better fill factors. This approach also allows the user to trade temporal resolution for imaging speed by reducing the number of time gates. For example, time gating was employed in SwissSPAD detectors with $512\times 512\text{pixels}$.³⁸ Therefore, as sensor sensitivity improves, parallel improvements in data streaming and analysis will be necessary to reach the full frame-rate potential of SPAD arrays for NAD(P)H light-sheet FLIM.

Overall, these NAD(P)H light-sheet FLIM systems will be an important new tool to study single cell metabolism and migration in 3D, including in vivo studies of whole model organisms.