Open Access
Add to BibSonomy
Abstract
Multiplexed fluorescence detection has become increasingly important in the fields of biosensing and bioimaging. Although a variety of excitation/detection optical designs and fluorescence unmixing schemes have been proposed to allow for multiplexed imaging, rapid and reliable differentiation and quantification of multiple fluorescent species at each imaging pixel is still challenging. Here we present a pulsed interleaved excitation spectral fluorescence lifetime microscopic (PIE-sFLIM) system that can simultaneously image six fluorescent tags in live cells in a single hyperspectral snapshot. Using an alternating pulsed laser excitation scheme at two different wavelengths and a synchronized 16-channel time-resolved spectral detector, our PIE-sFLIM system can effectively excite multiple fluorophores and collect their emission over a broad spectrum for analysis. Combining our system with the advanced live-cell labeling techniques and the lifetime/spectral phasor analysis, our PIE-sFLIM approach can well unmix the fluorescence of six fluorophores acquired in a single measurement, thus improving the imaging speed in live-specimen investigation.
© 2024 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
Traditional fluorescence imaging relies on emission filters to separate fluorescence signals of different colors. To achieve a better spectral separation result, researchers need to select fluorophores with minimal spectral overlap and use narrow-bandpass filters for multicolor imaging, which limits not only the dyes but also the photons available for investigating complex biological systems. Using the color-separation method alone, researchers can only differentiate three to four colors in the visible range [1,2]. To overcome this limitation, it is necessary to take advantage of other photophysical properties of the fluorophores in multiplexed imaging. One such property is the fluorescence lifetime. Being an intrinsic property of a fluorescent molecule, fluorescence lifetime measured by a FLIM system is not biased by excitation power or probe concentration [3]. In addition, lifetime reading is not prone to photobleaching [4] and can shed light on the microenvironment surrounding the fluorophore [5–7]. Moreover, fluorophores that share the same excitation/emission spectra can still be differentiated based on their distinct fluorescence lifetimes.
Several microscopic designs can provide simultaneous lifetime and spectral readings [7–12]. They often rely on a time-correlated single photon counting module [7] or a gated optical intensifier [11] for obtaining lifetime information, and a detector array [10] or multiple single-pixel detectors [12] for acquiring spectral information, creating what we call the first-generation spectral fluorescence lifetime microscopic (sFLIM) technologies for bioimaging. However, these first-generation sFLIM technologies had either limited spectral resolution, poor imaging quality or low imaging speed [12]. To obtain multiplexed FLIM images at high speed, Gratton’s group employed a digital-frequency-domain (DFD) lifetime acquisition module to analyze the signals from a photomultiplier tube (PMT) array for rapid and simultaneous spectral and lifetime measurements, creating what we call the second-generation sFLIM technologies that could blindly unmix the fluorescence from 3 dyes at each pixel using a phasor approach [13]. But the single excitation wavelength employed in the second-generation sFLIM approaches did not fully excite all 3 dyes, thus having a limited imaging quality due to the compromised signal-to-background ratio. In addition, although the second-generation sFLIM technologies offer unbiased multiplexed imaging, the current studies were restricted to fixed-cell demonstrations, due to the fact that the probes used were neither genetically encoded nor cell permeable.
Here we demonstrate a new multiplexed imaging method, which we call the third-generation sFLIM, that is designed specifically for live-cell imaging and can monitor up to 6 fluorophores simultaneously. In our system, we employ a pulsed interleaved excitation (PIE) scheme with two lasers to scan the live cells (Fig. 1(a)), a spectral fluorescence lifetime microscopic (sFLIM) system based on a 16-channel PMT array and an field programmable gate array (FPGA)-based, digital-frequency-domain (DFD [14,15]) acquisition unit to collect fluorescence (Fig. 1(b)), and a lifetime/spectral phasor algorithm [16,17] to blindly unmix the fluorescence at each pixel (Fig. 1(c)). Compared with the traditional color-separation method, our PIE-sFLIM method does not require samples to be scanned multiple times, thus greatly improving the imaging speed and lowering the phototoxicity. Compared with the second-generation technologies, our PIE-sFLIM method can unmix more dyes (from 3 dyes to 6 dyes) and achieve a 50% improvement in signal-to-noise ratio. Moreover, using self-labeling proteins (e.g., SNAP-tag [18] or HaloTag [19]) and cell-permeable dyes [20–22], multiplexed imaging of six cellular constituents in live T24 or HEK293 cells is demonstrated. This is the first report that combines live-cell labeling techniques with PIE and the phasor-sFLIM method in creating a new capability in live-cell multiplexed imaging.
Fig. 1. Schematic of the PIE-sFLIM system with an example of 16-channel raw data analysis. (a) PIE-sFLIM setup, where the two avalanche photodiode (APD) channels and the two-photon (2P) laser are for the traditional confocal and two-photon imaging. (b) 16-channel intensity with time-resolved fluorescence response (TRES) analysis for signal visualization, demonstrated on a convallaria sample with 488 nm excitation. (c) Representative decays curves for each spectral channel and spectrum at each pixel. The dips in the emission spectrum are due to the dichroic mirror that blocks the two potential excitation wavelengths at 560 nm and 650 nm. Abbreviations: 16-PMT, 16 channels photomultiplier tube; 2P, two-photon; FPGA, field programmable gate array; DFD, digital frequency domain; CFD, constant fraction discriminator; TRES, time-resolved emission spectrum; APD, avalanche photodiode; RL, relay lens; VP, variable pinhole; EM, emission filter; FL, focusing lens; IRB, infrared blocking filter; M, mirror; D, dichroic mirror; GSM, galvo mirror; SL, scan lens; Ex, excitation laser, Em, emission band.
2. Methods
2.1 Spectral fluorescence lifetime microscopy (sFLIM)
Our PIE-sFLIM is a method that combines true parallel, simultaneous lifetime, and spectral detection with phasor analysis to obtain fast, unbiased, high-precision FLIM data that can be processed in real time. It was built around a confocal and two-photon microscope (Eclipse Ti-S, Nikon) equipped with a motorized scanning stage for objective scanning (MS-2000XYZ, ASI). For excitation, two pulsed lasers emitting at 488 nm (denoted as Ex1, LDH-D-C-488, PicoQuant) and 650 nm (denoted as Ex2, SuperK EVO EU-4 supercontinuum laser, NKT Photonics) were employed. The lasers were operated in a pulsed interleaved excitation (PIE) mode, where the 488 nm laser was triggered by the 20 MHz sync signal from the supercontinuum laser. The 488 nm laser and the 650 nm laser were coupled to the laser scanning confocal system via two single mode polarization maintained (SMPM) optical fibers of different lengths. The length difference between the two optical fibers determined the delay of PIE, and the two PIE laser beams were finally combined in the laser scanning confocal system. To introduce a 25 ns delay between the 488 nm laser and the 650 nm laser (Fig. 2(a)), the length difference was set at 5 meters. The 488 nm light went through 2 meters SMPM optical fiber and the 650 nm light went through 7 meters SMPM optical fiber. To separate the excitation and the fluorescence signals, a custom-made dichroic mirror (zt473-491/561/640/2p-trans-pc, Chroma) was installed in the system.
Fig. 2. (a) PIE scheme and the resulting fluorescence decay measurement. Ex1 leads to fluorescence decay of Em1 and Em2 in Gate1, while Ex2 leads to fluorescence decay of Em3 in Gate2. (b) Normalized spectra of fluorescein and rhodamine B mixture, under 488 nm and 650 nm excitation, respectively, the ch9-12 (marked with the gray box) were saturated by the 650 nm excitation and thus not collected. (c) A 3D view of TRES using time-gated analysis. Em1- and Em2-Gate1 acquisition eliminates Em3 emissions (left white box L), while Em3-Gate2 acquisition eliminates Em1 and Em2 emissions (right white box R). (d) Temporally and spectrally resolved detection is divided into the three emission bands, where each band has a lifetime phasor plot that contains two of the six fluorophores used for live-cell imaging. Abbreviations: Ex, excitation laser; Em, emission band; Nuc, NucSpot488; Lyso, LysoTracker Green; Mito, MitoTracker Orange; SNAP, Lyn11-SNAP-SPY555-BG; HT9, Vimentin-HT9-JF646; HT7, β4Gal-HT7-JF646.
All images were acquired using a high-numerical, water-immersion objective (CFI60 60X/1.2 NA, Nikon). The fluorescence emitted by the sample was focused through an optical fiber bundle (Model A12420, Hamamatsu) onto a grating (A10766-Y016, Hamamatsu) and the dispersed light was collected by a 16-channel PMT array (H10515B-20, Hamamatsu). Having a resolution of 13.2 nm, the grating evenly covered a spectral range of 211.2 nm (from 500.7 nm to 711.9 nm) across the 16 channels. Each PMT channel was connected to a constant fraction discriminator (CFD) for precise timing of photon arrival. A FPGA-based DFD module (FastFLIM, ISS) then provided simultaneous fluorescence lifetime characterizations of the 16 channels (Fig. 1(b)) (Supplement 1 Note 1).
Each PIE-sFLIM image frame has 256 × 256 pixels with pixel dwell time of 50 µs, giving an image speed of 0.3 fps. Five frames are summed to ensure the brightest pixels in each emission band (Em1, Em2, and Em3) have at least 200 photon counts, corresponding to ∼50 photons per pixel per emission band on average. Therefore, the total acquisition time for generating one PIE-sFLIM image is 15 s. The acquisition parameters for the imaging experiments performed in this manuscript are listed in Supplement 1 Table S1. All acquired data were further analyzed using our customized PIE-sFLIM software (termed UT-PIE-sFLIM, an open-source, easy-to-use MATLAB software for phasor analysis and fluorescence unmixing).
In our PIE-sFLIM, lifetime information is sampled in 256 bins (from 0 to 50 ns), which creates an sFLIM dataset with the size of 256(x: pixels) × 256(y: pixels) × 256(t: time bins) × 16(λ: channels). This dataset, denoted as xytλ, is a multidimensional array where x and y represent the spatial dimensions (0.3 µm wide per pixel), t represents the fluorescence in a time bin (198 ps each bin), and λ represents the fluorescence in a color channel (13.2 nm wide per channel). At each pixel, there is a 256 × 16 sub-dataset termed the time-resolved emission spectrum [23] (TRES in Fig. 1(b)). Digital heterodyning performed by the DFD module turns the TRES into a cross-correlation phase histogram, which should represent the convolution of the fluorescence decay with the shape of the digital sampling [24]. From there, a phasor plot with multiple harmonic frequencies can be built, which is denoted as τ phasors (or lifetime phasors), as it contains the lifetime information of the embedded fluorophores. Similarly, the phasor obtained by transforming the spectral dimension is denoted as λ phasors (or spectral phasors). Our customized PIE-sFLIM unmixing software (termed UT-PIE-sFLIM in GitHub) takes τ and λ phasors as inputs, and outputs the weight of each fluorophore at each pixel. Using computer simulation and solution-based spike-in samples, the reliability and the accuracy of our unmixing algorithm was validated. We then tested our PIE-sFLIM imaging method on fixed and live cells labeled with diverse fluorescent dyes.
2.2 Pulsed interleaved excitation (PIE)
In laser scanning confocal microscopy, multiple continuous-wave (CW) lasers can be switched on and off by acousto-optic tunable filters (AOTF) in the microsecond scale to achieve alternating excitation in multiplexed imaging without crosstalk. This is sufficient for most cellular intensity imaging applications at the ensemble level. However, to achieve nanosecond temporal offset of two pulsed lasers, specifically for fluorescence lifetime multiplexed imaging, interlaced pulsed laser sources are needed. Alternating laser excitation (ALEX) was introduced by Kapanidis and Weiss [25,26]. They interleaved two excitation sources on a timescale between 25 and 3000 µs. By switching between both excitation sources on a timescale faster than diffusion of the particle through the probe volume, the labeling stoichiometry of individual complexes could be determined. Additionally, precise molecular sorting and detailed investigations into molecular interactions at the single-molecule level were also enabled by ALEX [27]. Later, with pulsed interleaved excitation (PIE) [28], Lamb’s group pushed the alternating excitation timescale down to the nanosecond regime. This fast interleaved excitation allowed dual-color fluorescence cross-correlation spectroscopy (FCCS) experiments to be performed with sub-microsecond resolution in probing the interactions between two molecules [29].
To effectively excite fluorophores with a wide range of colors, PIE based on two pulsed laser sources and a fiber-based temporal delay scheme [30,31] was employed in our system. The 488 nm light (denoted as Ex1) was from a laser diode (LDH-D-C-488, PicoQuant) whose emission could be triggered by the 20 MHz supercontinuum source emitting at 650 nm (Ex2, with an excitation filter 650/38 nm, Chroma). The timing precision of interleaved pulses was checked using the standard Ludox scatterer, a photodiode and an oscilloscope (Supplement 1). The FWHM of the pulses were 0.6 ns for Ex1 and 0.8 ns for Ex2, which served as the instrument response functions for lifetime calibration. The 16 spectral channels were divided into 3 emission color bands: green (Em1: ch1-4, 500.7 nm to 553.5 nm), orange (Em2: ch5-8, 553.5 nm to 606.3 nm), and red (Em3: ch13-16, 659.1 nm to 711.9 nm). Please note that the ch9-12 were saturated by the 650 nm excitation and thus not collected.
2.3 Cell culture and live-cell staining for PIE-sFLIM measurements
Either T24 or HEK293 cells were cultured in Dulbecco’s Modified Eagle Medium F12 (DMEM F12, #N6658-6X500 ML, Sigma-Aldrich) supplemented with 10% Fetal Bovine Serum (FBS, #10437028, Thermo Scientific) at 37°C in a 5% CO2 incubator. Cells were seeded on 3.5 cm glass bottom dishes (#NC1184521, MatTek). Multiple staining of the cells was done sequentially, for imaging lysosomes, nuclei, mitochondria, and plasma membranes in the Em1 and Em2 emission bands (Table 1). Dilutions of each dye were prepared in DMEM F12 medium. The stained cells were then transfected with two plasmids, β4Gal-HT7-Lyn11-Snapf (#175524, Addgene) and vimentin-HT9 (engineered in house), for imaging Golgi and vimentin in the Em3 band. The sequences of HT7 (HaloTag7) and HT9 (HaloTag9) can be found in literature [20]. Transient expression of the two plasmids was initiated by delivering 1 ug of each plasmid DNA using electroporation (Gene Pulser Xcell, Bio-Rad). Transfected cells were again seeded in 3.5 cm glass bottom dishes (#NC1184521, MatTek). After 24-hour incubation, cells were stained with 250 nM LysoTracker Green (#8783, Cell Signaling Technology) and 250 nM MitoTracker Orange (#M7510, Thermo Fisher Scientific) for 30 minutes, new DMEM F12 medium with 4000× diluted NucSpot488 (#40081-T, Biotium), 500 nM SPY555-BG (#CY-SC204, Cytoskeleton), and 40 nM JF646 (#6148, Bio-techne) for 1 hour in warm DMEM F12 medium in a cell-culture incubator. The excess dyes were removed by washing the cells three times using 1 mL warm PBS buffer (#10010023, Gibco). Cells were incubated for 1.5 hours before taking PIE-sFLIM measurements. Cell viability was checked by assessing the physical characteristics and appearance of cells to determine their health with epi-illumination microscopy. Details of cell viability assays are provided in Supplement 1 Note 4.
Table 1. Six fluorophores used to stain the live T24 and HEK293 cells for PIE-sFLIM imaging demonstration.
The fluorophores were chosen so that each emission band contained a pair of fluorophores that shared similar emission spectra (such as LysoTracker Green and NucSpot488 in the Em1 band) but had differentiable fluorescence lifetimes (3.8 vs. 5 ns). Although the 5th fluorophore and the 6th fluorophore in Em3 band were both HaloTag-dye complexes that shared the same JF646 dye, the distinct HaloTag constructs (HT7 vs. HT9 [20]) gave them distinguishable lifetimes. Thus, they were regarded as two different fluorophores. For the six fluorophores shown in Table 1, the Ex1 (488 nm) excited the four fluorophores in the Em1 and Em2 bands, while the Ex2 (650 nm) excited the two HaloTag-dye complexes in the Em3 band.
The complete excitation and detection scheme of the six fluorophores is shown in Fig. 2. The 25 ns pulse interval between Ex1 and Ex2 (ΔT) is long enough to ensure thorough fluorescence decay of the four Em1/Em2 fluorophores excited by Ex1 and thorough decay of the two Em3 fluorophores excited by Ex2 (Fig. 2(a)). Two commonly used dyes, fluorescein and rhodamine B, were mixed in solution for validation (Fig. 2(b)). Using Ex1 (488 nm), a small peak was seen at ch7 due to rhodamine B emission. Similarly, using Ex2 (650 nm), a low-level of fluorescein emission in ch4 was noted. To eliminate the bleed-through fluorescence, the 50 ns time-resolved acquisition period (i.e., the 256 time bins) was divided into two time gates, Gate1 (time bins 1-128) and Gate2 (time bins 129-256). Only the Em1 (ch1-4) from Gate1 following the 488 nm excitation was collected and used to represent fluorescein emission (thus bypassing the bleed-through at ch7), while only the Em2 (ch5-8) and Em3 (ch13-16) from Gate2 following the 650 nm excitation was used to represent rhodamine B emission (bypassing the bleed-through at ch4).
Similarly, in the case of 6 fluorophores, Em1 and Em2 fluorescence decays were only registered within Gate1 (0-25 ns, termed Em1-Gate1 and Em2-Gate1), while Em3 decay was only registered within Gate2 (25-50 ns, termed Em3-Gate2, Fig. 2(c)). This temporally and spectrally resolved measurement scheme guaranteed well differentiated lifetime phasors of the six fluorophores in the three emission bands (Fig. 2(d)).
2.4 Lifetime phasor analysis
For lifetime phasors (τ phasors), the multi-exponential fluorescence decay at each pixel in the xytλ dataset is transformed to a point (gτ, sτ) called a lifetime phasor based on Eq. (1) [32]:
To eliminate the artifacts given by the instrument response function and the delays of the electronics, phase $\varphi$ and modulation m of the phasor cloud [33] are first calibrated using well-characterized dyes, such as fluorescein (lifetime of 4 ns [14]) and Atto633 (3.3 ns, Fig. 3(b)). Figures 3(c), (d) show τ phasor calibration procedure based on fluorescein (for Em1 and Em2 bands) and Atto633 (for Em3 band), where m can be corrected by multiplying a constant factor α and $\varphi$ can be corrected by adding a constant offset Δ$\varphi$ to the initial value $\varphi$i, given by the following formulas:
Fig. 3. (a) Lifetime phasors can be generated using the measured phase delay φ(ω) and modulation m(ω) in a frequency-domain (FD) setup. For a single species, lifetime decreases clockwise along the universal semicircle. (b) Lifetime phasors can also be obtained from time-domain decay data, after conducting Fourier transform. The FLIM image shows T24 cells stained with LysoTracker Green (3.8 ns, green) and NucSpot488 (5 ns, blue). (c) Lifetime phasor calibration using fluorescein (4 ns, excited by 488 nm laser) for green (Em1) and orange (Em2) emission bands. (d) Lifetime phasor calibration using Atto633 (3.3 ns, excited by 650 nm laser) for red (Em3) emission band. (e) $\tau$ phasor plot of a convallaria sample. (f) False-colored FLIM image of the convallaria. Abbreviations: Nuc, NucSpot488; Lyso, LysoTracker Green. Scale bars are 5 µm.
Since the phasor cloud of a two-component mixture lies on a straight line joining the phasors of two individual components, the phasor cloud can be used to uncover the fractions of individual components at each pixel (Fig. 3(e), (f)).
For multiple fluorescent species, the relationships between phasors and lifetimes are given by Eq. (5):
2.5 Spectral phasor analysis
The analogy between lifetime and spectral measurements, along with the utilization of the phasor transform for spectral data instead of lifetime measurements, was first introduced by Gerritsen’s group [34]. For spectral phasor (λ phasors), the entire spectrum at each pixel in the xytλ dataset is transformed to a point (gλ, sλ) called a spectral phasor based on Eq. (7) [31]:
In addition to spectral phasor transform by post-processing as shown in our method, spectral phasor transform can also be conducted optically by using a pair of filters with sine/cosine transmission profiles [35]. Digman and Prescher have demonstrated live-cell multiplexed imaging with six bioluminescent reporters using this optical phasor transform approach [36].
The spectral phasor analysis shares many similarities with the lifetime phasor analysis, as both use the Fourier transform to turn the data (decay or spectrum) into a single phasor point in the phasor space defined by the real and imaginary parts of the transform. Comparable to lifetime phasors, pixels with similar spectral features form a phasor cloud in the λ phasor plot, where the spectral information is embedded in the phasor values. This feature allows interactive mapping between the pixels in an image and the corresponding λ phasors in a spectral phasor plot.
The 360° of the spectral phasor covers the range of 500.7-711.9 nm and is evenly divided by the 16 spectral windows (22.5° per window), with wavelength increasing in the counterclockwise direction.
To demonstrate the spectral phasor analysis, six Gaussian spectra with various FWHM (full-width-half-maxima) in the three emission bands were generated (Fig. 4(a)). Their corresponding λ phasors were largely determined by their emission peaks and FWHM, where phasors given by wide-spectrum emitters were closer to the center (0,0) and phasors given by red emitters were in the bottom half of the spectral phasor plot. As a result, λ phasors were clearly able to distinguish fluorophores shared an identical emission peak but with different emission profiles (such as fluorophore A: 579.9/30 nm, fluorophore B: 579.9/60 nm, and fluorophore C: 579.9/90 nm). When a large portion of the fluorescence emission was uncollected due to the limited detector range, fluorophores shared an identical emission peak could have their phasors not along the same radial line (such as fluorophore D: 527.1/50 nm and fluorophore E: 527.1/20 nm). This is because the DFT employed here works more effectively on a periodic function (i.e., a sine-like function vs. a step-like function) [31]. Nevertheless, there is no problem for using the emission spectrum features identified by the spectral phasors to differentiate fluorophores, as long as the same spectral acquisition settings are used.
Fig. 4. Spectral phasor plots of simulated Gaussian emission spectra centered at the peak wavelengths of 579.9 nm (A, B, C), 527.1 nm (D, E) and 670 nm (F) with the FWHM of A: 30 nm, B: 60 nm, C: 90 nm, D: 50 nm, E: 20 nm, F: 50 nm.
2.6 Hybrid unmixing
Here we present a new approach that leverages our PIE-sFLIM excitation/detection scheme and the Gaussian Mixture Model (GMM, see section 2.7) [37] to unmix fluorescence signals from multiple fluorophores based on their distinguished fluorescence lifetimes and emission spectra. The three-step workflow is described below, with a schematic representation shown in Fig. 5.
Fig. 5. (a) The TRES matrix I(t, λ) at each pixel is separately transformed into λ phasors and $\tau$ phasors. (b) λ phasor plots before and after spectral denoising. (c) $\tau$ phasor plots corresponding to the three emission bands. (d) Representative inputs and outputs of the Gaussian Mixture Models (GMM) based on unmixed spectra and lifetime phasors of the three emission bands. (e) Merged intensity image, merged multi-target image and unmixed images of T24 cells with nucleus stained with NucSpot488 (Nuc), lysosomes (Lyso) stained with LysoTracker Green, mitochondria (Mito) stained with MitoTracker Orange, plasma membrane (PM) stained with SPY555-BG (SNAP), Golgi stained with HT7-JF646 (HT7) and vimentin filament stained with HT9-JF646 (HT9). Scale bar is 20 µm.
(1) Transformation into Phasor Space: The 256 × 16 TRES matrix I(t, λ) at each pixel is separately transformed into λ phasors and $\tau$ phasors by Eq. (2) and Eq. (7) (Fig. 5(a)). An intensity threshold and 3 × 3 median filter are applied to the λ phasors and $\tau$ phasors to reduce the noise of the phasor location (Fig. 5(b)) [32,38]. When the image size is 256 × 256 pixels, this creates 256 × 256 sets of λ phasors (gλ, sλ) (Fig. 5(b)) and 256 × 256 × 3 sets of $\tau$ phasors (gτ, sτ) (Fig. 5(c)), considering that the lifetime phasor analysis is divided into the three emission bands. In other words, temporally and spectrally resolved fluorescence detection at each pixel eventually leads to one λ phasor set and three $\tau$ phasor sets (total of 4 phasor sets) at that pixel.
(2) Utilization of GMM for Pixel Clustering: The 256 × 256 phasor sets (gλ, sλ, gτem1, sτem1, gτem2, sτem2, gτem3, sτem3) are used as inputs for the GMM (Fig. 5(d)), with initial guesses on the number of species and their associated mean phasors. The GMM leverages its capability to understand data originating from multiple Gaussian distributions (representing distinct clusters) and learns their parameters such as mean, covariance, and weight (W, a probability of each pixel to belong to each of the clusters). Here GMM is used to classify each pixel and assign it to a corresponding cluster for color visualization on the lifetime phasor plot and spectral phasor plot (clustering). Note that this assignment is based on the dominant fluorescence signal at each pixel.
(3) Photon Redistribution and Unmixing: After the initial pixel classification, GMM is used again to provide a fully quantitative analysis of acquired fluorescence signals within each pixel, allowing us to weigh the photon counts in each emission band and redistribute them to each fluorophore cluster (unmixing). This redistribution process completes the PIE-sFLIM hybrid unmixing algorithm, generating unmixed images for individual fluorophores.
2.7 Gaussian mixture model
GMM is a probabilistic model widely used for clustering tasks [39]. It posits that observed data points originate from a combination of several Gaussian distributions, each representing a distinct cluster within the data. The GMM learns parameters such as mean, covariance, and weight for each Gaussian distribution, enabling the characterization of the underlying data distribution.
For our application, we adapt the GMM to unmix fluorescence signals in live-cell imaging based on both fluorescence lifetime and spectra. The GMM assumes that the observed fluorescence lifetime and spectra phasor plots can be represented as linear combinations of Gaussian components. GMM aims to estimate the parameters (means, covariances, and weights) that best describe the observed fluorescence lifetime and spectra distribution. The estimation is done using the Expectation-Maximization (EM) algorithm [40], which iteratively maximizes the likelihood of the observed data. Each Gaussian component within the GMM corresponds to a specific fluorophore in the sample. The mean of a Gaussian component signifies the fluorescence lifetime and the emission spectrum of the corresponding fluorophore, while the weight represents the proportion or abundance of that fluorophore in the mixture.
To build up the input dataset for our GMM, the TRES matrix I(t, λ) for each pixels is transformed into λ phasors and $\tau$ phasors to form a vector of 8 features: (gλ, sλ) - λ phasor, (gτem1, sτem1) - green (Em1) $\tau$ phasor, (gτem2, sτem2) - Orange (Em2) $\tau$ phasor, (gτem3, sτem3) - red (Em3) $\tau$ phasor. The complete output dataset is a table with 7 columns and 256 × 256 rows, as shown in the top row of Fig. 5(d).
Estimating the number of distinct fluorophores in a sample is critically challenging. To address this issue, we establish an automated approach that determines the optimal number of clusters using the Bayesian Information Criterion (BIC) [41,42]. By fitting the GMM with varying numbers of clusters, we compute BIC values and select the model that best balances goodness of fit and model complexity. The approach works best for blind unmixing of unknown fluorescent species, which is the case in our convallaria imaging experiment (section 3.1). This automatic estimation of cluster number gives more flexibility to our approach.
In certain scenarios where prior knowledge of fluorophores in use is available, such as the six fluorophores used in the live T24 and HEK293 experiments, the GMM is given an initial number of clusters and their associated mean λ phasors (gλ, sλ) and mean τ phasors (gτ, sτ). Specifically, in the case of six fluorophores, the number of clusters is pre-defined as six, and we establish a distinct set of mean λ (gλ, sλ) and τ (gτ, sτ) phasors for each individual fluorophore. In the six fluorophores scenario, we set, NucSpot488 being cluster_1 assigned with index 1 (IDX: 1), LysoTracker Green being cluster_2 (IDX: 2), MitoTracker Orange being cluster_3 (IDX: 3), SPY555-BG being cluster_4 (IDX: 4), and HT7-JF646 being cluster_5 (IDX: 5) and HT9-JF646 being cluster_6 (IDX: 6). For each cluster, the mean λ phasor (gλ, sλ) and mean τ phasors (gτ, sτ) are calculated from the measured λ phasors (Fig. 8) and τ phasors of the pure species (Fig. 2(d)). Temporally and spectrally resolved detection is divided into the three emission bands, where each band has a lifetime phasor plot that contains two of the six fluorophores used for live-cell imaging. Examples of the optional initial cluster values for cluster_1 (NucSpot488), cluster_3 (MitoTracker Orange), and cluster_5 (HT7-JF646) are shown in the middle row of Fig. 5(d), while a completed list of the initial values for all six clusters is provided in Supplement 1 Table S2. Our approach of initializing GMM clusters’ parameters allows for a more guided unmixing process, resulting in enhanced accuracy and efficiency, particularly when the number of fluorophores is known a priori.
After defining the inputs, our GMM is fitted with an Expectation-Maximization (EM) algorithm [43,44]. The EM algorithm iteratively estimates the parameters of the Gaussian components and assigns phasor points to clusters based on the estimated probabilities. This process continues until convergence, and the resulting model outputs the unmixing results at each pixel which are then used for generating the unmixed and merged multi-target images. The outputs are formatted into a table as shown in the bottom row of Fig. 5(d). Each pixel has a weight for each of the clusters/fluorophores, where the largest weight determines the index (IDX) of that pixel. For pixel_1, the highest weight 0.7 is at cluster_1 (Wg1, NucSpot488). Thus, pixel_1 has the index of 1, which is used to color code pixel_1 (cyan) on the unmixed λ and $\tau$ phasor plots (Fig. 5(d) bottom and Fig. 9(b)). The weights are then used to redistribute the photon counts collected at this pixel to each fluorophore, thus generating unmixed images (Fig. 5(e) and Fig. 9(d)). Assuming that 800, 100 and 100 photons are collected at pixel_1 in Em1-Gate1, Em2-Gate1 and Em3-Gate2, respectively, 700, 100, 50, 50, 50 and 50 photons will be assigned to cluster_1 through cluster_6, accordingly.
The unmixed clusters and the weight distribution of data points within those clusters are visualized through unmixed λ and $\tau$ phasor plots, $\tau$ phasor contour plots [39], and unmixed intensity image for each fluorophore (Fig. 5(d), (e)). This comprehensive visualization allows us to intuitively assess the quality of unmixing and the accuracy of fluorophore identification. The Gaussian Mixture Models Unmixing pseudocode is provided in Supplement 1 Note 2. By leveraging the power of GMM, we effectively separate and quantify different fluorophores within a mixed sample. The automatic estimation of the number of fluorophores, along with the option for user-initiated seed-based estimation, contributes to the versatility of our method.
3. Results
3.1 Convallaria samples
Figures 6 and 7 show the measurements and processed results of the convallaria sample (#MSAS0321, IDS) using our PIE-sFLIM system. The merged image is produced by summing all TRES photons at each pixel. The number of fluorescent species is determined by BIC and tested with our GMM algorithm. The λ phasor plot clearly identifies three fluorescent species in the convallaria sample (Fig. 6(a)), where the short-wavelength species are colored in green and yellow, while the long-wavelength species is pseudo-colored in red (Fig. 6(b)).
Fig. 6. (a) The spectral phasor unmixing of the convallaria sample with 3 species denoted as cluster 1 - green, cluster 2 - orange, cluster 3 - red. (b) Unmixed intensity images of cluster 1 - green, cluster 2 - orange, cluster 3 - red and the merge intensity image. Scale bar is 10 µm.
Fig. 7. The lifetime phasor unmixing of the convallaria sample with 3 species denoted as cluster 1 - green, cluster 2 - orange, cluster 3 - red, with lifetime image and overlay with intensity image corresponding to (a) green emission band (Em1) with Gate1, (b) orange emission band (Em2) with Gate1, (d) red emission band (Em3) with Gate1, and (c) red emission band (Em3) with Gate2. Scale bar is 10 µm.
Different from the single λ phasor plot, there are three τ phasor plots, which correspond to Em1-Gate1, Em2-Gate1 and Em3-Gate2 acquisitions (Fig. 2(a), (c)). Since the three fluorescent species identified in the λ phasor plot also have different fluorescence lifetimes, three fluorescent species can be seen in all three τ phasor plots (Fig. 7(a)-(c)). The staining dyes in the convallaria sample are well separated spectrally and temporally, therefore spectral phasor unmixing and lifetime phasor unmixing yield the same number of clusters and a similar imaging result. The fluorophore patterns shown in the lifetime unmixed images are similar to that in the spectrally unmixed images, where the species of a short lifetime has a long emission wavelength and vice versa. To compare the performance of our PIE-sFLIM system with that of the second-generation sFLIM system, which uses only a single laser for excitation, we plot theτ phasor based on Em3-Gate1 acquisition under only the 488 nm laser excitation (Fig. 7(d)). The 488 nm laser (Ex1) clearly does not fully excite all 3 species, leading to low photon counts in Em3-Gate1 acquisition. This further results in a noisy τ phasor plot (Fig. 7(d) left and Fig. S3) and a low-quality intensity image (Fig. 7(d) right). In contrast, the 650 nm laser (Ex2) well excites the red species, nearly doubling the acquired photon counts (Supplement 1 Fig. S3) in Em3-Gate1 and thus improving both the τ phasor plot (Fig. 7(c) left) and the resulting imaging quality (Fig. 7(c) right). The signal-to-noise ratio (SNR) of a τ phasor can be evaluated based on the spread along the direction of the phasor elongation [15,45]. We adopted the same criterion to assess the SNR of the PIE-sFLIM generated phasors. The direction of elongation of a given τ phasor was identified using linear regression. The spread perpendicular to the elongation direction, ΔN, represented the SNR, where a smaller ΔN meant a higher SNR. We then calculated ΔN for each pixel on the phasor plot. The average ΔN served as a metric to compare the SNR of the phasor plot among different methods. Following this criterion, the SNR improvement by using PIE-sFLIM approach in the multi-target imaging, as compared to the single-excitation approach [13], is estimated to be 50% (Supplement 1 Note 3).
3.2 Live-cell samples
For sFLIM imaging, fluorophore preference is given to those with single-exponential decays and distinguishable fluorescence lifetimes. In our experiment, there is fluorescence lifetime difference of at least 0.5 ns among the chosen pairs of probes (Table 1). Other than using commercially available live-cell dyes to stain organelles such as nucleus (NucSpot 488), lysosomes (LysoTracker Green) and mitochondria (MitoTracker Orange), we also take advantage of the self-labeling protein (SLP) tags to stain other cellular constituents. SLP tags, such as HaloTag7 [19], HaloTag9 [20,21] and SNAP-tag [18], react with fluorophores bearing a chloroalkane (CA) and benzylguanine (BG) ligand, respectively. The use of SLP tags allows us to image ER, Golgi and vimentin - thanks to the recent progress in engineering SLP variants for fluorescence lifetime multiplexing [20].
Cells were stained with six fluorophores following the protocol described in section 2.3. Preliminary rounds of imaging were done with cells incubated with a single dye in order to characterize its lifetime spread in live cells (Table 1). Subsequent rounds of incubation were performed, each gradually increasing the number of fluorophores in the sample. This allowed us to adjust the incubation time and dye concentration such that the fluorescence intensity of each of the probes was in the same order of magnitude.
For initial screening purposes, we chose to assess the fluorescence lifetime properties based on τ phasor analysis as it allowed rapid visualization screening of probes with large differences in fluorescence lifetime without the need for fitting. The differences in fluorescence lifetime within one emission band was assessed by overlaying the measured phasors of the pure species (Fig. 2(d)). The largest difference in fluorescence lifetimes was found in the green emission band (Em1) between LysoTracker Green (3.8-4 ns) and NucSpot488 (5-5.2 ns), followed by MitoTracker Orange (2-2.2 ns) and Lyn11-SNAP-SPY555-BG (2.5-2.8 ns) in the orange emission band (Em2). The probes in the red emission band (Em3) had the smallest lifetime difference between β4Gal-HT7-JF646 (3.5-3.8 ns) and Vimentin-HT9-JF646 (4-4.4 ns).
Spectral phasor analysis was also used for rapid visualization screening of probes with distinct fluorescence emission spectrum signatures (Fig. 8). Although HT7-JF646 and HT9- JF646 were considered as two different fluorophores due to their different fluorescence lifetimes, they shared the same emission spectrum. As discussed in section 2.5 and Fig. 4, emission peak revealed by the phasor plot could be skewed when fluorophore’s emission spectrum is largely outside the detector range (e.g., NucSpot488, LysoTracker Green and JF646). Nevertheless, as long as the five dyes are well resolved on the λ phasor plot, the unmixing algorithm can be smoothly executed.
Fig. 8. (a) Emission spectra and (b) spectral phasors of the five dyes (NucSpot488, LysoTracker Green, MitoTracker Orange, SPY555-BG, and JF646) used for PIE-sFLIM demonstration in live T24 and HEK293 cells. Abbreviations: Lyso, LysoTracker Green, Nuc, NucSpot488; Mito, MitoTracker Orange; SNAP, Lyn11-SNAP-SPY555-BG; JF646, Vimentin-HT9-JF646 and β4Gal-HT7-JF646.
Live T24 (Fig. 5(d)) and HEK293 (Fig. 9) cells were imaged using PIE-sFLIM and the resulting fluorescence was recorded over the 16 spectral channels grouped into three emission bands (Fig. 2(d)). The irradiation power of the lasers varied depending on the labeling efficiencies but remained within a similar range. Among the six fluorophores, it was the most challenging to differentiate HT7-JF646 and HT9-JF646. But the structures of Golgi and vimentin were still well resolved in the unmixed images. Note that the hybrid unmixing procedures achieved in Fig. 5(e) and Fig. 9(d) are not possible to obtain using emission filters, since the emission peaks of the two fluorophores in each emission band differ by as little as 5 nm. Moreover, the median execution time for running the unmixing algorithm, based on 10 cells, is under 50 ms, meaning that unmixing can be conducted in nearly real time following data acquisition at a pixel. In contrast to other blind unmixing methods that rely on phasors, we only exploit the first harmonic of the phasor transformation, which is the least affected by noise. We noticed some cytosolic signals were seen in the HEK293 nucleus image, but not in the T24 nucleus image. As for the presence of cytosolic signals in nuclei, it was indeed an issue in imaging live HEK293 cells, where the mitochondrial DNA also seemed to be stained by NucSpot488. But for live T24 cells, NucSpot488 signals were confined to nuclei, as shown in Fig. 5(e).
Fig. 9. (a) Unmixed spectral phasor plot for six fluorophores (NucSpot488, LysoTracker Green, MitoTracker Orange, SPY555-BG and JF646). (b) Unmixed lifetime phasor plot for green, orange and red emission bands. (c) Contour lifetime phasor plot for green, orange and red emission bands. (d) Merge image and unmixed FLIM images of HEK293 cells with nucleus stained with NucSpot488 (Nuc), lysosomes stained with LysoTracker Green (Lyso), mitochondria stained with MitoTracker Orange (Mito), plasma membrane (PM) stained with SPY555-BG (SNAP), Golgi stained with HT7-JF646 (HT7) and vimentin filament stained with HT9-JF646 (HT9). Scale bar is 20 µm.
4. Discussion
In this study, we introduced the third-generation sFLIM, termed Pulsed Interleaved Excitation Spectral Fluorescence Lifetime Microscopy (PIE-sFLIM), for simultaneous imaging of six fluorescent species in live cells. Although multiplexed cellular imaging based on PIE and lifetime/spectral readings was previously reported by Sauer and coworkers [46], their demonstration was limited to fixed cells. Besides, their pattern-matching algorithm used for fluorescence contribution analysis required precise reference patterns of fluorescence decay and emission spectrum signatures obtained from cell samples labeled with individual dyes. In contrast, our PIE-sFLIM method relies on the phasor analysis and a “blind” unmixing scheme that does not require any prior knowledge on the fluorescence decay and spectrum signatures of individual fluorophores. Using the phasor analysis and a confocal setup with four excitation/emission channels and FastFLIM system, Frei and Johnsson have achieved imaging of eight probes in live U2OS cells, where each of the four emission channels contained two probes with distinct lifetimes [21]. However, without using the PIE scheme, the emission and lifetime measurements in the four color channels need to be acquired separately (i.e., scan the sample four times), thus not suitable for probing any fast dynamic processes in live cells.
Our PIE-sFLIM system is different, as it inherits the advantages of both the PIE scheme [46] (simultaneous data acquisition with high excitation efficiency) and the second-generation sFLIM methods [13] (DFD-based rapid lifetime acquisition and blind unmixing of three dyes). In other words, PIE-sFLIM not only requires only a single scan for multiplexed imaging but also performs fast lifetime measurements at each pixel. In addition, using two excitation lasers, our PIE-sFLIM system fully excites the six fluorophores used for live-cell imaging, doubling the acquired photon counts when compared to the single-excitation sFLIM approach [13]. A 50% improvement in the SNR has been observed in both the τ phasor plots (Fig. 7(c) left) and the final multi-target images (Fig. 7(c) right, as compared to Fig. 7(d) that is acquired by the previous single-excitation sFLIM method).
Our PIE-sFLIM approach has overcome the limitations of traditional color-separation methods and previous sFLIM techniques, providing enhanced imaging speed, unbiased unmixing of fluorescence signals, and compatibility with advanced live-cell labeling techniques [47–49]. This advancement has the potential to significantly impact various fields, including biosensing and bioimaging, by enabling the investigation of rapid dynamic cellular events with improved spatial and temporal resolution.
Funding
University of Texas at Austin (Texas Global Faculty Research Seed Grant); National Science Foundation (CBET 2235455); National Eye Institute (EY033106).
Disclosures
The authors declare no conflicts of interest.
Data availability
Data underlying the results presented in this paper are not publicly available at this time but may be obtained from the authors upon reasonable request.
Supplemental document
See Supplement 1 for supporting content.
References
1. D. Lin, X. Li, E. Moult, et al., “Time-tagged ticker tapes for intracellular recordings,” Nat. Biotechnol. 41(5), 631–639 (2023). [CrossRef]
2. L. Wang, M. Tran, E. D’Este, et al., “A general strategy to develop cell permeable and fluorogenic probes for multicolour nanoscopy,” Nat. Chem. 12(2), 165–172 (2020). [CrossRef]
3. K. Suhling, L. M. Hirvonen, J. A. Levitt, et al., “Fluorescence lifetime imaging (FLIM): Basic concepts and some recent developments,” Medical Photonics 27, 3–40 (2015). [CrossRef]
4. M. Y. Berezin and S. Achilefu, “Fluorescence Lifetime Measurements and Biological Imaging,” Chem. Rev. 110(5), 2641–2684 (2010). [CrossRef]
5. H. Wallrabe and A. Periasamy, “Imaging protein molecules using FRET and FLIM microscopy,” Curr. Opin. Biotechnol. 16(1), 19–27 (2005). [CrossRef]
6. W. Schrimpf, J. Jiang, Z. Ji, et al., “Chemical diversity in a metal–organic framework revealed by fluorescence lifetime imaging,” Nat. Commun. 9(1), 1647 (2018). [CrossRef]
7. W. Becker, “Fluorescence lifetime imaging – techniques and applications,” J. Microsc. 247(2), 119–136 (2012). [CrossRef]
8. P. Tinnefeld, D.-P. Herten, M. Sauer, et al., “Photophysical Dynamics of Single Molecules Studied by Spectrally-Resolved Fluorescence Lifetime Imaging Microscopy (SFLIM),” J. Phys. Chem. A 105(34), 7989–8003 (2001). [CrossRef]
9. J.-P. Knemeyer, D.-P. Herten, M. Sauer, et al., “Detection and Identification of Single Molecules in Living Cells Using Spectrally Resolved Fluorescence Lifetime Imaging Microscopy,” Anal. Chem. 75(9), 2147–2153 (2003). [CrossRef]
10. F. Fereidouni, K. Reitsma, H. C. Gerritsen, et al., “High speed multispectral fluorescence lifetime imaging,” Opt. Express 21(10), 11769–11782 (2013). [CrossRef]
11. D. M. Owen, E. Auksorius, H. B. Manning, et al., “Excitation-resolved hyperspectral fluorescence lifetime imaging using a UV-extended supercontinuum source,” Opt. Lett. 32(23), 3408–3410 (2007). [CrossRef]
12. R. Borlinghaus and L. Kuschel, “Spectral fluorescence lifetime imaging microscopy: new dimensions with Leica TCS SP5,” Nat. Methods 3(10), 868 (2006). [CrossRef]
13. L. Scipioni, A. Rossetta, G. Tedeschi, et al., “Phasor S-FLIM: a new paradigm for fast and robust spectral fluorescence lifetime imaging,” Nat. Methods 18(5), 542–550 (2021). [CrossRef]
14. Y.-I. Chen, S.-C. Liao, T. D. Nguyen, et al., “Generative adversarial network enables rapid and robust fluorescence lifetime image analysis in live cells,” Commun. Biol. 5(1), 18 (2022). [CrossRef]
15. Y.-I. Chen, Y.-J. Chang, Y. Sun, et al., “Spatial resolution enhancement in photon-starved STED imaging using deep learning-based fluorescence lifetime analysis,” Nanoscale 15(21), 9449–9456 (2023). [CrossRef]
16. Q. S. Hanley, “Spectrally resolved fluorescent lifetime imaging,” J. R. Soc. Interface. 6, S83–S92 (2008). [CrossRef]
17. Y. Zhou, J. M. Dickenson, Q. S. Hanley, et al., “Imaging lifetime and anisotropy spectra in the frequency domain,” J. Microsc. 234(1), 80–88 (2009). [CrossRef]
18. A. Keppler, S. Gendreizig, T. Gronemeyer, et al., “A general method for the covalent labeling of fusion proteins with small molecules in vivo,” Nat. Biotechnol. 21(1), 86–89 (2003). [CrossRef]
19. G. V. Los, L. P. Encell, M. G. McDougall, et al., “HaloTag: A Novel Protein Labeling Technology for Cell Imaging and Protein Analysis,” ACS Chem. Biol. 3(6), 373–382 (2008). [CrossRef]
20. M. S. Frei, M. Tarnawski, M. Julia Roberti, et al., “Engineered HaloTag variants for fluorescence lifetime multiplexing,” Nat. Methods 19(1), 65–70 (2022). [CrossRef]
21. M. S. Frei, B. Koch, J. Hiblot, et al., “Live-Cell Fluorescence Lifetime Multiplexing Using Synthetic Fluorescent Probes,” ACS Chem. Biol. 17(6), 1321–1327 (2022). [CrossRef]
22. J. B. Grimm, B. P. English, J. Chen, et al., “A general method to improve fluorophores for live-cell and single-molecule microscopy,” Nat. Methods 12(3), 244–250 (2015). [CrossRef]
23. J. H. Easter, R. P. DeToma, L. Brand, et al., “Nanosecond time-resolved emission spectroscopy of a fluorescence probe adsorbed to L-alpha-egg lecithin vesicles,” Biophys. J. 16(6), 571–583 (1976). [CrossRef]
24. R. A. Colyer, C. Lee, E. Gratton, et al., “A novel fluorescence lifetime imaging system that optimizes photon efficiency,” Microsc. Res. Tech. 71(3), 201–213 (2008). [CrossRef]
25. A. N. Kapanidis, N. K. Lee, T. A. Laurence, et al., “Fluorescence-aided molecule sorting: Analysis of structure and interactions by alternating-laser excitation of single molecules,” Proc. Natl. Acad. Sci. 101(24), 8936–8941 (2004). [CrossRef]
26. A. N. Kapanidis, T. A. Laurence, N. K. Lee, et al., “Alternating-Laser Excitation of Single Molecules,” Acc. Chem. Res. 38(7), 523–533 (2005). [CrossRef]
27. N. K. Lee, A. N. Kapanidis, Y. Wang, et al., “Accurate FRET Measurements within Single Diffusing Biomolecules Using Alternating-Laser Excitation,” Biophys. J. 88(4), 2939–2953 (2005). [CrossRef]
28. B. K. Müller, E. Zaychikov, C. Bräuchle, et al., “Pulsed Interleaved Excitation,” Biophys. J. 89(5), 3508–3522 (2005). [CrossRef]
29. J. Hendrix, W. Schrimpf, M. Höller, et al., “Pulsed Interleaved Excitation Fluctuation Imaging,” Biophys. J. 105(4), 848–861 (2013). [CrossRef]
30. Y. Sun, U. Coskun, A. C. M. Ferreon, et al., “Tunable PIE and synchronized gating detections by FastFLIM for quantitative microscopy measurements of fast dynamics of single molecules,” SPIE Proceedings (2016).
31. Y. Sun, T. D Nguyen, Y.-I. Chen, et al., “Characterization of a time-resolved spectral detector for spectral fluorescence lifetime imaging with the parallel 16-channel FastFLIM and the phasor analysis,” SPIE Proceedings (2023).
32. M. A. Digman, V. R. Caiolfa, M. Zamai, et al., “The Phasor Approach to Fluorescence Lifetime Imaging Analysis,” Biophys. J. 94(2), L14–L16 (2008). [CrossRef]
33. S. Ranjit, L. Malacrida, D. M. Jameson, et al., “Fit-free analysis of fluorescence lifetime imaging data using the phasor approach,” Nat. Protoc. 13(9), 1979–2004 (2018). [CrossRef]
34. F. Fereidouni, A. N. Bader, H. C. Gerritsen, et al., “Spectral phasor analysis allows rapid and reliable unmixing of fluorescence microscopy spectral images,” Opt. Express 20(12), 12729–12741 (2012). [CrossRef]
35. P. N. Hedde, R. Cinco, L. Malacrida, et al., “Phasor-based hyperspectral snapshot microscopy allows fast imaging of live, three-dimensional tissues for biomedical applications,” Commun. Biol. 4(1), 721 (2021). [CrossRef]
36. Z. Yao, C. K. Brennan, L. Scipioni, et al., “Multiplexed bioluminescence microscopy via phasor analysis,” Nat. Methods 19(7), 893–898 (2022). [CrossRef]
37. C. M. Bishop, Pattern Recognition and Machine Learning (Springer, 2006), p. 738.
38. H. J. Chiang, D. E. S. Koo, M. Kitano, et al., “HyU: Hybrid Unmixing for longitudinal in vivo imaging of low signal-to-noise fluorescence,” Nat. Methods 20(2), 248–258 (2023). [CrossRef]
39. A. Vallmitjana, B. Torrado, E. Gratton, et al., “Phasor-based image segmentation: machine learning clustering techniques,” Biomed. Opt. Express 12(6), 3410–3422 (2021). [CrossRef]
40. T. Huang, H. Peng, K. Zhang, et al., “Model Selection for Gaussian Mixture Models,” STAT SINICA (2017).
41. G. Bianchetti, M. De Spirito, G. Maulucci, et al., “Unsupervised clustering of multiparametric fluorescent images extends the spectrum of detectable cell membrane phases with sub-micrometric resolution,” Biomed. Opt. Express 11(10), 5728–5744 (2020). [CrossRef]
42. S. Gideon, “Estimating the Dimension of a Model,” Ann. Statist. 6(2), 461–464 (1978). [CrossRef]
43. A. P. Dempster, N. M. Laird, D. B. Rubin, et al., “Maximum Likelihood from Incomplete Data via the EM Algorithm,” Journal of the Royal Statistical Society. Series B (Methodological) 39(1), 1–22 (1977). [CrossRef]
44. S. K. Ng, T. Krishnan, G. J. McLachlan, et al., Handbook of Computational Statistics: Concepts and Methods (Springer, 2012), pp. 139–172.
45. M. J. Sarmento, M. Oneto, S. Pelicci, et al., “Exploiting the tunability of stimulated emission depletion microscopy for super-resolution imaging of nuclear structures,” Nat. Commun. 9(1), 3415 (2018). [CrossRef]
46. T. Niehörster, A. Löschberger, I. Gregor, et al., “Multi-target spectrally resolved fluorescence lifetime imaging microscopy,” Nat. Methods 13(3), 257–262 (2016). [CrossRef]
47. J. Kompa, J. Bruins, M. Glogger, et al., “Exchangeable HaloTag Ligands for Super-Resolution Fluorescence Microscopy,” J. Am. Chem. Soc. 145(5), 3075–3083 (2023). [CrossRef]
48. L. Hellweg, A. Edenhofer, L. Barck, et al., “A general method for the development of multicolor biosensors with large dynamic ranges,” Nat. Chem. Biol. 19(9), 1147–1157 (2023). [CrossRef]
49. T. D. Nguyen, Y. I. Chen, L. H. Chen, et al., “Recent Advances in Single-Molecule Tracking and Imaging Techniques,” Annu. Rev. Anal. Chem. 16, 253–284 (2023). [CrossRef]
