NIR Fluorescence lifetime macroscopic imaging with a novel time-gated SPAD camera

Abstract

SwissSPAD3 is the latest of a family of widefield time-gated SPAD imagers developed for fluorescence lifetime imaging (FLI) applications. Its distinctive features are (i) the ability to define shorter gates than its predecessors (width W < 1 ns), (ii) support for laser repetition rates up to at least 80 MHz and (iii) a dual-gate architecture providing an effective duty cycle of 100%. We present widefield macroscopic FLI measurements of short lifetime NIR dyes, analyzed using the phasor approach. The results are compared with those previously obtained with SwissSPAD2 and to theoretical predictions.

1. INTRODUCTION

Fluorescence lifetime imaging provides a wealth of information on a sample of interest, that is complementary to the intensity recorded at each pixel of an image^1-3. In microscopy, this type of measurement is generally performed using a confocal geometry in which a focused pulsed laser excitation is raster-scanned over the sample, and the fluorescence signal recorded by a point detector (photomultiplier – PMT, hybrid photodetector – HPD or single-photon avalanche diode – SPAD). When the sample size is macroscopic, a different approach is generally used, based on widefield excitation with either a modulated or a pulsed laser source, detection of the fluorescence signal being performed with a modulated or time-gated intensified charge-coupled detector (ICCD). The cost of these setups can be significant and their complexity, together with the fragility and limited sensitivities of ICCDs, justifies investigating alternative approaches. CMOS SPAD imagers offering matrices of individually readout SPADs have reached maturity during the past decades, with several commercial devices now available⁴. By design, they enable parallel acquisition of data in each pixel of the image (such as ICCDs), providing in principle higher throughput than raster scanning approaches.

Two main approaches have been developed for these devices: a time-correlated time-tagged approach and a time-gated one. In the time-correlated time-tagged approach (also referred to as time-correlated single-photon counting – TCSPC), each photon detected in every SPAD is timed precisely with respect to the laser pulse. This is the most photon efficient approach, as all photons are registered (up to a hardware specific maximum count rate), but also technically more challenging, as complex electronics is needed to achieve precise timing for every photon in every pixel. It is currently limited to SPAD arrays of fairly modest size (~ 200 × 200), although progress in the technology will likely remove this limitation. Additionally, it results in very high data transfer throughput unless some data processing and/or compression is performed before transfer to the host computer.

By contrast, the SwissSPAD series of detectors^{5, 6}, of which the detector discussed in this work, SwissSPAD3 (SS3)⁷, is the latest generation, uses a time-gating principle. In this approach, each SPAD is programmed to be sensitive to incoming photons only during a predefined temporal window (known as the gate, defined by its width W and its offset t_i with respect to the laser pulse) during each laser period. In effect, the detector is blind during the period outside this window, and photon collection efficiency is therefore reduced. Moreover, in order to provide information about the fluorescence decay, the gate needs to be repeatedly shifted across the whole laser period, which results in sequential acquisition of the temporal profile. This can in principle lead to artifacts if the sample’s intensity varies during acquisition. Additionally, because of this sequential acquisition, the more gate steps G are used to acquire a complete decay (i.e. the higher the decay sampling resolution), the longer the acquisition.

In previous works with the second generation of SwissSPAD detectors (SS2)⁸, we have studied the interplay between these parameters (gate width W and number of gate number G) and shown that in a first approximation, the detector’s performance for fluorescence lifetime extraction is bias-free and shot-noise limited (Eq. (1)):

$$\frac{{\sigma }_{\tau }}{\tau }=\frac{F\left(W,G,\tau \right)}{\sqrt{N}}$$ (1)

where τ is the lifetime and σ_τ its standard deviation (or uncertainty), N the number of detected photons used for lifetime estimation, and F a factor depending on the gate width W, gate number G and lifetime τ. Importantly, this relationship demonstrates that there is no limit to the achievable lifetime precision, no matter how large the gate width is, provided the number of accumulated photons N is sufficiently large.

In a later study of SS2’s performance of short lifetime extraction at high laser repetition rate, we showed that factor F is very sensitive to the gate width W, and in particular diverges when W → kT, where T is the laser period and k an integer⁹. This divergence is due to the fact that a gate width equal to the laser period results in a signal that does not depend on the actual gate offset t_i, since the gate is always opened, i.e. a constant signal, from which no lifetime information can be extracted.

SwissSPAD3⁷ was designed to improve on SS2 in several aspects. First, the minimal gate width was reduced from ~10 ns (for SS2) to below 1 ns. Second, the imager’s response uniformity was improved and its rising and falling edge characteristic time scales reduced. Finally, the ability to record photons impinging on each SPAD outside of the gate window was added to increase the effective duty cycle to 100%.

This article reports a preliminary investigation of the effect of these modifications on lifetime precision as obtained with samples of near-infrared (NIR) dye samples relevant for in vivo imaging.

2. METHODS

2.1. SS3 detector

SS3 has been extensively described in Arin Ülkü’s Ph D thesis¹⁰ and in ref. ⁷. Briefly, the device is comprised of the CMOS SPAD chip (500 × 500 SPADs) wire-bonded to a first PCB equipped with a C-mount adapter to attach an imaging lens. This PCB is connected to a larger one equipped with input/output connectors to provide power to the electronics, receive and send digital signals (including the laser trigger signal). To that larger PCB are connected one (or two) smaller one(s) equipped with a field-programmable gate array (FPGA) to control the device’s function and receive/transmit data to the host computer using either a USB3 connector or a PCIe connector (only the USB3 link was used in this work). A single FPGA board was used in this work, which allowed to control and readout only half of the sensor’s SPADs, resulting in 500 × 250 pixel images.

2.2. Optical Setup

The macroscopic fluorescence lifetime imaging (MFLI) setup used in this work was identical to that described previously⁹. Briefly, a pulsed laser emitting sub 100 ps pulses at 765 nm with an 80 MHz repetition rate (VIS-IR-765, Picoquant, USA), was coupled to a single-mode fiber and recollimated, expanded and homogenized with an engineered diffuser (ED1-S20-MD, Thorlabs) to form a homogeneous square illumination pattern on the sample plane. Excitation intensity was adjusted by inserting a neutral density filter before the diffuser. Fluorescence light emitted by the sample was collected with a NIR macroscopic lens attached to the SS3 detector, after rejection of the laser wavelength by a bandpass filter (FF01-893/209-25, Semrock, NY). The excitation and emission paths were enclosed in a black box preventing NIR light to escape and ambient room light to be detected by the instrument.

Instrument response functions (IRFs) were acquired with the same setup without the emission filter (and a stronger neutral density filter) using a sheet of white filter paper to scatter the laser light.

2.3. Samples

Two standard NIR dyes, ICG (IR-125, Exciton, OH) and IRDye 800CW (929-70020, Li-Cor) were used as provided by the manufacturer and diluted in dimethyl sulfoxide (DMSO). ICG concentrations ranging from ~10 μM to ~10 nM and IRDye 800CW (abbreviated IR800 in the following) concentrations ranging from ~1 μM to ~1 nM were prepared and deposited in a black 96 well-plate. The wells were sealed with a transparent adhesive sheet to prevent evaporation during the measurements.

IRFs were acquired in identical conditions as the samples’ fluorescence signal (except for the neutral density filter and emission filter) using a piece of white filter paper as scattering sample placed at the same location as the fluorescence sample. One IRF dataset and background file was acquired for each gate width settings.

Detector background files were acquired in identical conditions as for the samples, by blocking the laser source.

2.4. Data Acquisition

Data was recorded using an updated version of SwissSPAD Live, a custom software written in Lab VIEW (NI, Austin, TX)¹¹. The software provides a user-friendly interface to define gate width W, gate number G, step size δt and integration time t_int, loads the FPGA firmware controlling the sensor and receives the transferred data, which is then saved in an open-source file format based on the HDF5 hierarchical data format after proper decoding. In contrast to SS2, SS3 data is comprised of to images per gate position: the gate image, consisting in the number of photons detected during the opened gate duration (similar to SS2 data), and the intensity image, recording all photons detected during the frame duration. In this work, the gate opening and closing sequence was kept active during the whole frame duration set to 40 μs. Gate steps of δt = 18.6 ps (the minimum step size when running SS3 at 80 MHz) were used, requiring G = 672 gates to cover the laser period T = 12.5 ns. Gate width W ranging from 10 ns down to below 1 ns were used. 10-bit images were recorded. The intensity image data was saved but not used.

2.5. Data Analysis

Data analysis was performed using AlliGator¹², a free software developed for fluorescence lifetime analysis, as described for SS2, with the exception that the additional information contained in the intensity channel was used to enhance signal-to-noise ratio. Briefly, after pile-up correction, detector background was subtracted in each channel and normalization of the gate channel by the intensity channel was performed using the following equation:

$$I_{{}_{G,corr}}\left(t_{{}_i}\right)=I_{{}_G}\left(t_{{}_i}\right){}^{\ast }\frac{\max (I_{{}_{INT}})}{I_{{}_{INT}}\left(t_{{}_i}\right)}$$ (2)

where t_i is the gate offset with respect to the laser pulse and I_G denotes the signal recorded in the gate channel and I_INT the signal recorded in the intensity channel.

Regions of interest (ROIs) were defined for the brightest dataset by specifying a background threshold and used for all datasets (see Fig. 1A). Individual pixels of each ROI were analyzed individually as described next.

Fig. 1:

Data analysis example. A: Intensity image for a 1 ns gate width. B: top left ICG solution ROI decay (black), corresponding IRF (blue), single-exponential fit (red) and residuals (green). The fitted lifetime is τ_ICG = 0.72 ns. C: Calibrated phasor plot for the dataset shown in A. The two clusters corresponding to the ICG wells (phase lifetime ~ 0.7 ns) and IRDye800CW wells (phase lifetime ~ 1 ns) are visible.

The software allows plotting the decay corresponding to each pixel (or if needed each ROI) or compute their discrete phasor¹², using the formula:

$$z\left[\left\{F_{{}_j}\right\}\right]=g+is=\frac{\sum _{j=1}^G{F}_{{}_j}{e}^{i2\pi f{t}_j}}{\sum _{j=1}^G{F}_{{}_j}}$$ (3)

where F_j is the pixel intensity in gate image j (1 ≤ j ≤ G), with gate offset t_j = t₀ + (j-1)δt, f is the phasor harmonic frequency (f = 1/T in this work) and i denotes the imaginary square root of −1. (g, s) are the phasor components (real and imaginary part, respectively).

Gate widths were determined by fitting the recorded temporal profile of laser light scattered by a piece of white filter paper using a logistic square gate function, as described previously¹³.

Fluorescence decays were fitted by a single exponential function convolved with the IRF corresponding to that pixel using a non-linear least square fit routine based on the Levenberg-Marquardt algorithm (see Fig, 1B for an example of full ROI decay and fit). In this work, a single-exponential model was fitted, from which the fitted lifetime τ is obtained. Phasors were calibrated¹² using the corresponding IRF phasor (defined by τ_IRF = 0) and represented in a standard phasor plot (see Fig. 1C for an example). The corresponding phase lifetime τ_φ was computed according to the standard equation, valid for large gate numbers (G = 672 in this work)¹⁴:

$${\tau }_{{}_{\phi }}=\frac{1}{2\pi }\frac{s}{g}$$ (4)

For each ROI, the standard deviation of all fitted lifetimes (resp. all phase lifetimes) was computed and is noted σ_τ (resp. σ_τφ). Data discussed in this article correspond to the phase lifetime τ_φ and its corresponding standard deviation σ_τφ.

3. RESULTS AND DISCUSSION

3.1. Phase lifetime dispersion as a function of gate width W

A series of measurements of a dilution series (~10 μM to ~10 nM) of the NIR dye ICG and of the NIR dye IRDye800 CW (~1 μM to ~1 nM) in dimethyl sulfoxide (DMSO) was performed by varying the gate width parameter W from 10 ns (the maximum achievable) to 0.4 ns. The mean and standard deviation in each well, as well as a box representing the 25 to 75 percentiles of the phase lifetime distribution, are represented for W = 0.5, 1, 5 and 10 ns in Fig. 2.

Fig. 2:

Phase lifetime distributions as a function of gate width W. Examples of phase lifetime box plots (25%-75% box, mean ± standard deviation) are represented for 4 different gate width W (0.5, 1, 5 and 10 ns). In each plot, the ICG wells data are on the left, ordered in decreasing concentration, and the IRDye 800CW wells on the right. The plot index n (in ICGn and IR800n) corresponds to the well number in Fig. 1 (numbered from left to right).

The first noteworthy feature of these plots is the concentration dependence of both dyes lifetimes, in particular for IRDye800CW: the lifetime decreases from ~1 ns at 1 μM down to ~0.6 ns at ~3.3 nM. A similar, yet less dramatic change is observable for ICG, whose lifetime is ~0.7 ns from 10 μM to 100 nM, and decreases down to ~0.55 ns at 10 nM. Concentration effects on lifetime are well known in dye solutions and usually interpreted as due to aggregation and the formation of multimers resulting in self-quenching, and are therefore important to be aware of when using NIR dyes for in vivo measurements¹⁵. Importantly, this feature is reproduced in all datasets, irrespective of the gate width W.

The second striking feature is the increasing spread of phase lifetime distributions as concentration decreases for each dye (from left to right). This has a simple explanation, since the recorded signal is proportional to the concentration (see next section). The concentration ratio between successive wells of a given dye being 1/3, the total drop in concentration for the observed ICG wells is (1/3)⁶ = (1/27)², and (1/3)⁵ = (1/15.6)² for IRdye800 CW, hence an expected drop of SNR (or increase of standard deviation) by a factor 27 for ICG and 15.6 for IRDye800CW, between the most and least concentrated samples. A closer analysis of this relation is discussed in the next section.

Finally, it is also clear that the dispersion of the phase lifetime for a given sample depends on the gate width W: it is notably less spread out for larger gates (W = 5 and 10 ns) than it is for shorter ones (W = 0.5 and 1 ns) in the examples shown in Fig. 2.

3.2. Shot-noise influence on phase lifetime precision

As noted previously, aside from the gate width W changing from one measurement to the next, two related effects are at play in each well measurement: the dye concentration is different (potentially resulting in quenching and thus, lifetime and signal reduction) and consequently, the recorded signal level is different, decreasing with concentration. SwissSPAD3 being a photon-counting device, it is therefore interesting to investigate whether the observed phase lifetime dispersion can be interpreted as being due to shot-noise only, or whether some additional factors might be at play. This section will present a preliminary analysis of this question.

Fig. 3 represents various observables and derived quantities as a function of gate width for the brightest wells of each dye solution series (leftmost wells in Fig. 1). The mean phase lifetime τ_φ represented in Fig. 3A shows no significant variation across all gate widths W. The phase lifetime standard deviation σ_τφ represented in Fig. 3B however, indicates a noticeable increase at short gate width for both dyes, as well as some localized increase around W = 8 ns IRDye800CW. Interestingly, their ratio, the relative phase lifetime uncertainty represented in Fig. 3C, appears almost identical.

Fig. 3:

Analysis of shot-noise and gate width influence on phase lifetime precision. Plots of mean phase lifetime τ (A), phase lifetime standard deviation σ (B), relative phase lifetime uncertainty σ/τ (C), mean photon count <N> (D), measured F-factor (E) and Fisher information result for F (F) are represented as a function of W for the most concentrated well of each dye sample (ICG: black, IRDye 800CW: red). The relative lifetime τ/T and relative width W/T allow using the theoretical results for arbitrary laser period T (in these experiments, T = 12. 5 ns).

The total photon count N per well in the gate channel (after normalization by the intensity channel, which in the present case does not modify the total signal significantly, as little gate offset dependence of the intensity signal was observed) shown in Fig. 3D shows a linear increase of signal as a function of gate width as expected, since the effective exposure is proportional to the gate width and acquisition was performed in a signal regime where negligible detector saturation occurs.

Fig. 3E shows the value of the F-factor (or photon economy) defined by:

$$F=\frac{\sigma }{\tau }\sqrt{N}$$ (5)

which indicates the departure of the method to the ideal statistical case F = 1 corresponding for instance to a detector recording each photon’s arrival time accurately and a method extracting the lifetime optimally.

F(W) increases approximately linearly within the range of gate width 0.4 ns – 10 ns, with F(0) = 4 for ICG (τ = 0.72 ns) and F(0) = 2.8 for IRDye800CW (τ = 1 ns) and similar slopes of 0.79 and 0.66 ns⁻¹ respectively. This behavior departs from the theoretically expected behavior represented in Fig. 3F for different values of the lifetime τ in the range [0.625 – 1.13] ns, which predicts a decreasing slope as lifetime increases but an asymptotic value F(0) = 1 in all cases, and a sharp divergence as W → T. This difference between theory and observation suggests that a number of important detector and experimental specificities are not taken into account by the simple model used to compute the Fisher information-based result of Fig. 3A. In particular, the effect of background subtraction and gate channel normalization could possibly degrade the expected performance. Further studies will be needed to fully understand the origin of this discrepancy.

Regardless of the reason for the observed dependence of F on τ and W, the observed approximate linear dependence of F on W:

$$F(\tau ,W)\sim F_{{}_0}(\tau )+\lambda (\tau )W$$ (6)

and the linear dependence of the signal N on width W (for a given sample and integration time t):

$$N(W)=k(t)W$$ (7)

yields the following empirical relation for the relative lifetime precision:

$$\frac{\sigma }{\tau }=\frac{F}{\sqrt{N}}\sim \frac{1}{\sqrt{k(t)}}\left(\frac{F_{{}_0}(\tau )}{\sqrt{W}}+\lambda (\tau )\sqrt{W}\right)$$ (8)

where W is expressed in ns and k and λ are unitless. Note that this heuristic formula does not capture the expected divergence of σ/τ as W → T, but indicates the existence of an optimal gate width W* given by:

$$W^{{}^{\ast }}=\sqrt{\frac{\lambda (\tau )}{F_{{}_0}(\tau )}}$$ (9)

As indicated by Fig. 3C, this minimum is fairly flat and slightly different gate widths will generally provide similar lifetime resolution.

4. CONCLUSION AND PERSPECTIVES

We have described preliminary measurements of time-gated NIR fluorescence decays performed with SwissSPAD3, a novel time-gated SPAD camera with a dual-gate architecture, capable of achieving gate widths from 10 ns down to less than 0.5 ns. We used a series of solutions of ICG or IRDye800CW in DMSO, whose lifetime was measured as a function of gate width W. The measured lifetimes were independent of W, demonstrating the ability of SS3 to perform equally well with short or long gates.

While the samples studied here were not displaying any short- or long-term intensity variations, the ability to compensate for such variations by normalization of the gate channel by the intensity channel, which records the total signal during the whole individual frame exposure, should allow acquiring reliable fluorescence lifetime data even in the presence of slow varying signal, for instance due to bleaching. The typical time scale of these variations should however be longer than the duration of a single gate image acquisition (about 40 ms in the present measurements) for this correction to be effective.

An investigation of the influence of signal intensity, which scales linearly with the gate width, on the precision of the lifetime extracted using the phasor approach, showed that, for a given integration time, short gate widths result in lower lifetime precision. While it appears that an optimal gate width exists for best relative lifetime precision, SS3 measurements of short lifetime NIR dyes reported here indicate reliable measurements no matter what gate width is used, the performance only noticeably decreasing for short gate duration.

With its ability to natively function with high frequency laser (such as the 80 MHz pulsed laser used in the experiments reported here), better response homogeneity over the whole detector area (data not shown), as well as narrower rise and fall times (data not shown), and dual-gate architecture, SS3 represents a significant improvement over its predecessor. By design, the SwissSPAD detector series is capable of advanced data processing as well as various data acquisition schemes thanks to its flexible FPGA firmware. Future work will be focused on taking advantage of these capabilities to address challenging practical fluorescence lifetime imaging scenarios.