Using in vivo multiphoton fluorescence lifetime imaging to unravel disease-specific changes in the liver redox state

Multiphoton microscopy with fluorescence lifetime imaging (MP-FLIM) enables deep, specific, quantitative and safe imaging of various organ sub-architecture and redox processes in vivo. We have now used MP-FLIM to study the pathophysiology and xenobiotic transport in a range of organs, including the liver, the kidney and the skin [1–17]. In this work, we have responded to requests to revisit previous work that examined changes in NAD(P)H and collagen using MP-FLIM with a focus on the liver to expand upon the analysis presented.

The analysis is important in the context of the liver being the key organ in the synthesis, metabolism and excretion of key endogenous and exogenous solutes, including being the key detoxifying organ for all solutes absorbed from the gastrointestinal tract. In particular, the liver processes carbohydrates, fats and proteins from the digestive tract to produce energy, metabolises and excretes into the bile various endogenous substrates, xenobiotics and their metabolites and can facilitate enterohepatic recirculation [18]. Hence, the liver plays a major role in ultimately regulating what is circulated to the rest of the body. We have shown that various liver diseases can compromise the liver's metabolic function and that the extent of impairment may be directly related to the severity of the liver disease and the nature of the solutes involved [19–21]. In these early studies, we used destructive sampling to assess the fate of different drugs in liver disease. In later studies, we began non-invasively imaging fluorescent drugs and their metabolites in the liver [2, 22]

It is now well known that the redox state and reactions of the liver are critical to metabolic and other functions of the liver and their modification by exogenous compounds and disease can impair liver function [23]. Whilst the autofluorescent intracellular coenzymes, reduced nicotinamide adenine dinucleotide (NADH) and flavin adenine dinucleotide (FAD), are key contributors to cell respiration and energy homeostasis, the autofluorescent reduced nicotinamide adenine dinucleotide phosphate (NADPH) is a key liver redox regulator involved in reductive biosynthesis and a cofactor involved in the glutathione scavenging of free radicals, heme-mediated P450 enzyme oxidation of many solutes [24, 25] and the flavin-containing monooxygenases (FMO) 2, 3, 4 and 5 catalyze the oxidation of various endogenous and exogenous solutes with amine, sulphur, and phosphorous groups [26]. Importantly, the autofluorescent fully oxidized (quinone form) coenzyme, FAD, which is reduced to its hydroquinone form, FADH₂, is also involved in this later process. The reaction rate for NADH is usually several orders lower than that of NADPH [27]. However, NADH is a key source of NADPH and is generated by several enzymes in both the liver cytoplasm and mitochondria. The oxidised forms of NAD⁺ and NADP⁺, as well as the reduced FADH₂, are not fluorescent. Glock and McLean [28] have reported NADP⁺, NADPH, NAD⁺ and NADH concentrations (μg g⁻¹ tissue) of 425, 116, 25 and 231, respectively. Given that NADPH and NADH have identical fluorescence signatures and cannot be separated in MP-FLIM imaging [29], the two entities are collectively referred to here as NAD(P)H.

The metabolic cofactors NAD(P)H and FAD can also exist in both an unbound (free) form and in a protein bound form, with these forms having different fluorescence lifetimes. Many studies have used MP-FLIM to distinguish between the free and bound forms of NAD(P)H and FAD in various cell studies [30–32] but, to our knowledge, studies of alterations in these coenzymes in liver disease have been limited to recent conference presentations [33, 34]. In general, differences in emission wavelength bands are used to separate NAD(P)H from FAD whereas differences in fluorescence lifetimes are used to separate their free and bound forms and are usually derived using time-correlated single-photon counting. In the latter, a double exponential decay function, f(t) (equation (1)) is used to model photon counts generated over time in an emission window after a laser pulse at a particular excitation wavelength:

$$\begin{eqnarray}&&f\left(t\right)=\,{a}_{1}{e}^{-t/\tau 1}+{a}_{2}{e}^{-t/\tau 2}+C\,{\rm{with}}\,{a}_{1}+{a}_{2}=1\end{eqnarray} \tag{ 1 }$$

where a₁, a₂ are the fluorescence amplitudes and τ₁, τ₂ are the fluorescence lifetimes of the short and long decay components. The instrument response function (IRF) of each FLIM image is convolved with the model function, f(t), in order to obtain the function F(t). The function F(t) is fitted to each measured pixel to calculate decay parameters. For NAD(P)H, the short decay components correspond to free NAD(P)H and the long decay components correspond to bound NAD(P)H. In contrast, the short decay components for FAD correspond to bound FAD and the long decay components to free FAD. However, recent high-resolution in vitro measurements of adipocytes showed that in a fluorescence lifetime analysis based on two exponential model fitting gives rise to two components that are coupled to the mitochondria while the other two components are coupled and arise from the cytoplasm [35]. In addition, in vivo measurements of the brain have been more accurately modeled with 4 exponentials rather than 2, where the first two components relate to free NAD(P)H [36].

In our previous work, we used morphology and amplitude mean fluorescence life time (τ_m,a) of NAD(P)H to understand the effects of liver pathophysiology on liver function [9]. Here, we undertook a more complete intravital MP-FLIM analysis of the liver redox state examining changes not only in free and bound NAD(P)H but also free and bound FAD in normal mice and in mice with different liver pathologies: ischemia-reperfusion injury, advanced liver fibrosis induced by carbon tetrachloride (CCl₄), primary sclerosing cholangitis in a genetic knock out mouse model (Mdr2-/-), fatty liver disease, and hepatocellular carcinoma (HCC). We used the second harmonic generation fluorescence lifetime associated with collagen to define fibrotic collagen deposition in CCl₄ induced fibrotic liver, primary sclerosing cholangitis, fatty liver disease and hepatocellular carcinoma. In order to more fully understand the impact of liver disease on the redox processes in the liver, we compared the results of two of the common algorithms used in FLIM analyses, undertook more sophisticated multiexponential fitting of our photon count - time data and used pixel segmentation and binning methods to improve the robustness of our analyses. We found with more detailed analyses that fibrosis, either with the CCl₄ or Mdr2-/- model produced significant changes in NAD(P)H and collagen and that changes could be detected with all three segmentation methods and with all three metabolic ratios: redox ratio, NAD(P)H amplitude ratio, and fluorescence lifetime redox ratio (FLIRR). We also found that the three metabolic ratios and (τ_m,a) were useful in demonstrating metabolic differences between control mice and those with hepatocellular carcinoma.

Liver disease models

Liver imaging

In this work, we reanalyzed the MP-FLIM liver images reported previously [9], focusing mainly on cellular metabolic alteration in the hepatocytes in the capsule of the liver, but also on hepatic stellate cells and Kupffer cells, within the structural context of hepatic sinusoids and biliary structures (figure 1(a)). We sought to measure changes in the fluorescence lifetimes of autofluorescent free and bound NAD(P)H, as well as the autofluorescent free and bound FAD, to characterise changes in the liver redox microenvironment, which is at the nexus of many metabolic pathways (figure 1(b)). MP-FLIM was performed using the DermaInspect system (Jen-Lab GmbH, Jena, Germany) with a 80-MHz titanium sapphire laser, laser power set to 15 mW (MaiTai, Spectra Physics, Mount View, California) and imaging with an oil-immersion 40 X objective (Carl Zeiss, Germany). FLIM data was acquired using a time-correlated single-photon counting SPC-830 detector (Becker & Hickl, Berlin, Germany), using a range of two photon (2P) excitation wavelengths (λ_ex) and fluorescence emission wavelength bands (λ_em). Figure 1(c) shows the autofluorescence images from our previous work [9] acquired at λ_ex 740 nm and λ_em 350 to 450 nm that show the range of morphological changes induced by the different liver pathologies. Other autofluorescence images were taken at λ_ex 800 nm with λ_em bands of 350 to 450 nm and 515 to 620 nm. Figure 1(d) shows the interrelationship between these ranges and the excitation and emission spectra for NAD(P)H, collagen and FAD. Also shown in figure 1(d) are the emission bands used by Cao et al [41] in their simultaneous imaging of NAD(P)H and FAD at 800 nm.

Zoom In Zoom Out Reset image size

Mean fluorescence lifetime analysis

We estimated both the mean amplitude-weighted (τ_m,a) and intensity-weighted (τ_m,i) lifetimes for NAD(P)H and FAD using equations (2) and (3). The commonly used τ_m,a is more biased to short fluorescent lifetime components and better defines relative changes in all components, including relative changes in NAD(P)H and FAD ratios for bound and unbound moieties [42]. In contrast, τ_m,i is dominated by the longer fluorescence lifetimes component and more closely approximates an assumed single-exponential approximated lifetime [42].

$$\begin{eqnarray}&&{\tau }_{m,a}=\displaystyle \frac{{a}_{1}{\tau }_{1}+{a}_{2}{\tau }_{2}}{{a}_{1}+{a}_{2}}\end{eqnarray} \tag{ 2 }$$

$$\begin{eqnarray}&&{\tau }_{m,i}=\displaystyle \frac{{a}_{1}{{\tau }_{1}}^{2}+{a}_{2}{{\tau }_{2}}^{2}}{{a}_{1}{\tau }_{1}+{a}_{2}{\tau }_{2}}\end{eqnarray} \tag{ 3 }$$

Metabolic ratios

We used three metabolic ratios in this analysis: (i) the optical redox ratio (Redox), (ii) the NAD(P)H amplitude ratio (NAD(P)H ratio), and (iii) the fluorescence lifetime redox ratio (FLIRR).

Redox measures the extent to which NAD(P)H and FAD endogenous fluorescence is quenched after liver injury and is defined as:

$$\begin{eqnarray}&&Redox=\displaystyle \frac{{I}_{FAD}}{{I}_{NAD\left(P\right)H}\,}\end{eqnarray} \tag{ 4 }$$

Where I_FAD is the intensity of FAD signal and I_NAD(P)H is the intensity of NAD(P)H.

The NAD(P)H ratio defines the relationship between unbound to bound NAD(P)H and is defined as:

$$\begin{eqnarray}&&NAD\left(P\right)H\,ratio=\displaystyle \frac{{a}_{1}}{{a}_{2}}\end{eqnarray} \tag{ 5 }$$

Where a₁, a₂ are the fluorescence amplitudes of the short (unbound) and long (bound) decay components of NAD(P)H.

FLIRR is the ratio of the fractional amplitudes of the bound decay components of NAD(P)H and FAD and is defined as:

$$\begin{eqnarray}&&FLIRR=\displaystyle \frac{{a}_{2{\rm{NAD}}({\rm{P}}){\rm{H}}}}{{a}_{1{\rm{FAD}}}\,}\end{eqnarray} \tag{ 6 }$$

Where a_2NAD(P)H is the fluorescence amplitude of the long (bound) decay component of NAD(P)H and a_1FAD, is the fluorescence amplitude of the short (bound) decay component of FAD.

FLIM model analysis

Segmentation methods

Statistical analysis

Initially, we used both FLIMFit and SPCImage 5.5 to process our imaging data and examined various bi- and tri- exponential model non-linear regression models. However, we found that FLIMFit allowed more control over the fitting, flexibility and reproducibility with segmentation, enabled maximum likelihood-based regression and produced more usable visualization of statistics than SPCImage 5.5.

Effect of binning on model parameters

The heterogeneity of NAD(P)H expression in the liver disease models prompted us to analyze the effect of binning on the fitted data. We found that spatial binning, commonly known as binning factor, which is (2n + 1)² where n is the number of pixels, in which the photons are accumulated from multiple adjacent pixels to create one data point, can compromise spatial resolution of in vivo liver images. This difficulty is probably specific for in vivo and ex vivo imaging of organs because by their nature, body organs are heterogeneous, consisting of a number of different cells along with intercellular and vascular spaces. In contrast, binning should increase the confidence of parameters derived from a homogeneous image, be it a solution, a solid or a monolayer of a specific cell culture. As our next step, we then assessed binning factors that may be optimal to define the various parameters for our in vivo heterogeneous organ system. We found that, when we compared binning factors of 0, 1, 3 and 6, which are defined by no binning, binning of 3 × 3 pixels, 7 × 7 pixels and 13 × 13 pixels of data, respectively, there were significant differences in the parameter estimates of a_1, τ₁ and τ₂ between binning factors 0, 1, 3 and 6 in most of the liver disease models (figures 2(a)–(c)). In particular, the Mdr2-/- fibrosis resulted in aberrant averaging of the data across cells and bile ducts (figures 2(a)–(c)). The spatial binning 1 (3 × 3 pixels) proved to be the optimal trade-off between having sufficient accumulated photons for robust model fitting (figure 2(d)) and the spatial resolution required to discern differences between the different cells and spaces.

Zoom In Zoom Out Reset image size

Effect of segmentation on model parameters

As the intensity images available for each of the individual disease states did not evenly fill the field of view, image analyses were undertaken based on regions of interest. As both a binning and a segmentation operation was performed on each image, we examined whether the order of operation could affect the data analysis. We found that the processes of (1) segmenting the data first and then binning it, with model fitting in FLIMFit, versus (2) binning the data, with model fitting in FLIMFit, and segmenting the resultant data led to a significant difference in the parameter values generated (figure 3(a)). The main reason for this difference is that, in (1) segmentation leads to the pixels outside the mask being set to zero and included in the model fitting and the binned data, with a decreased number of photons available for model fitting relative to (2) in which segmentation occurs last. We therefore used method (2) in our subsequent analyses, i.e. we binned data with factor 1, performed bi-exponential model fitting to the whole image and then segmented the sub regions of the image for further analysis.

Zoom In Zoom Out Reset image size

In our heterogeneity analysis of the different cell types and spaces in the liver, we investigated the model parameters in the following different segments of the image: whole region, cytoplasm, liver sinuses, and nucleus (figure 3(b)). In general, similar a₁ values were found for the different segmented regions of the NAD(P)H channel data in the normal liver. However, the Mdr2-/- liver sinuses had lower a₁ values (figure 3(c)) compared to elsewhere and in the normal livers. But, these values did not translate through to the parametric maps of NAD(P)H a₁/a₂ ratio, where we were unable to visually discern differences (figure 3(d)), although quantification of the pixels highlighted the difference in the Mdr2-/- sinuses (figure 3(e)).

A key impact on our observations was the choice of the instrument response function (IRF). Our observation was that the IRF generated using the FLIMFit methodology from a reference image gave us longer lifetime results for our normal liver image data, as for a₁ in figure 4 condition 1. However, when we imported the IRF derived by SPCImage 5.5 into FLIMFit, we obtained bi-exponential parameters more consistent with the published values (figure 4 conditions 2, 3, 4, 5). Unfortunately, IRF alone did not reduce the large confidence intervals for a₁ (figure 4 condition 2), which we showed earlier could be addressed by increased binning. But, binning itself is a 'double-edge sword'. Whilst binning can improve parameter precision, it does so at the expense of optimal spatial resolution (figure 4 condition 3). Thus, our key challenge in in vivo imaging remains having sufficient binning to get parameter precision but not so much as to lose the details in the liver heterogeneity. Overall, we found that segmenting an image after the model fitting (figure 4 condition 4), yielded more consistent and meaningful outcomes compared to segmenting the image before a model fit (figure 4 condition 5), by preserving the heterogeneity information.

Zoom In Zoom Out Reset image size

Finally, we examined whether the type of non-linear regression algorithm and number of exponents used in the regressions affected outcomes. Here, we used SPCImage 8 to compare analysis of our images with the maximum likelihood estimate and weighted least squares fit methods. We found the latter resulted in decreased and less reliable a₁ values (figure 4 conditions 6, 7). Consequently, all of our subsequent results used maximum likelihood based regressions and were conducted, because of its ease of use, with FLIMFit. We also used bi-exponential nonlinear regressions in all subsequent work as we did not find any significant improvement in using a tri-exponential regression based on chi-squared (data not shown). These findings were confirmed based on the model selection criteria for non-linear regressions based on two, three and four exponential decays using the programme Scientist (MicroMath Scientific Software, Salt Lake City, UT).

Average amplitude weighted and intensity weighted fluorescence lifetimes

Average amplitude weighted and intensity weighted fluorescence lifetimes for various liver diseases were compared to normal (figure 5). It is evident that our τ_m,a for the whole region segmentation using FLIMFit are similar to those we had previously reported based on analysis with SPCImage 5.5 using a bi- exponential decay model [9]. In particular, our analysis confirmed that the τ_m.a of NAD(P)H for all liver diseases were significantly different from the τ_m.a for normal livers (figure 5(a)). Our analysis yielded τ_m,a values for NAD(P)H (mean ± stdev) of: normal liver, 1834 ± 205 ps; ischemia-reperfusion injury at 6 h, 1694 ± 204 ps; ischemia-reperfusion injury at 24 h, 1776 ± 236 ps; CCl₄ induced liver fibrosis, 1918 ± 319 ps; Mdr2-/- livers, 1978 ± 292 ps; fatty livers, 1625 ± 246 ps; and HCC, 1707 ± 307 ps. Our data showed similar results to those previously reported for NAD(P)H in cholestasis and cirrhosis [33, 34].

Zoom In Zoom Out Reset image size

Our results also shows that the choice of segmentation method yielded similar estimates for both the NAD(P)H τ_m,a and NAD(P)H τ_m,i of all models (figure 5(a)). The similarity in the results for each of the segmentation methods probably reflects the dominance of the NAD(P)H signal in the hepatic cells in determining the τ_m,a and τ_m,i values. In particular, the FLIMFit program segmentation algorithm otsu_oht produces segmentation maps that are comprised of pixels from within cells in the NAD(P)H channel as well as the pixels in between the cells, as shown in figure 1(e). It is therefore more inclusive than manual cell segmentation but similar to whole region segmentation. The overall result is then a similar mean and range in τ_m,a and τ_m,i derived (figure 5(a)). However, it is also evident τ_m,a values for the different liver states vary more than τ_m,i for the same livers for the automation segmentation analysis. It is to be noted that, in general, the τ_m,i for all livers and segmentation methods is much greater than the τ_m,a, consistent with the greater weighting τ₂ has on τ_m,i relative to τ_m,a [42].

Figure 5(b) shows our corresponding MP-FLIM findings for the λ_ex 800 nm and λ_em 350–450 nm channel. It is evident that the τ_m,a results in Mdr2-/-, fatty liver and HCC were significantly lower than in normal livers, regardless of the segmentation method. Interestingly, the τ_m,i results parallel those for τ_m,a in showing a significant decrease only in HCC livers, whereas τ_m,i results from Mdr2-/- mice increased relative to normal livers (figure 5(b)).

Analysis of the distinct FAD fluorescence lifetimes are shown for various segmentations. Our initial analysis of the segmented FAD channel (λ_ex 800 nm; λ_em 515–650 nm) data resulted in a₁ values that spanned the whole range from 0 to 1, suggesting that no confidence could be placed in the mean a₁ value. We therefore only included pixels with a minimum intensity of 100 in our subsequent regressions. We found that when an intensity threshold was applied to the data, there was a significant decrease in the number of usable pixels in each image, leading to limited usable data after cell segmentation (figure 5(c)), and that the fibrotic livers had the most remaining pixels that could be analyzed. Both τ_m,a and τ_m,i increased in the CCl₄ induced and Mdr2-/- fibrosis models derived using automated and whole region segmentation methods. The values obtained for FAD τ_m,a (mean ± stdev) were: normal liver, 1185 ± 707 ps; ischemia-reperfusion injury at 6 h, 1691 ± 819 ps; ischemia-reperfusion injury at 24 h, 1028 ± 734 ps, CCl₄ induced liver fibrosis, 1155 ± 529 ps; Mdr2-/- mice livers, 1259 ± 410 ps; fatty liver disease was 1083 ± 831 ps; and HCC, 834 ± 394 ps.

NAD(P)H model individual fitted parameters

Figure 6 shows the individual parameters derived from the bi-exponential regressions of the MP-FLIM data for NAD(P)H in different liver diseases relative to control livers and for different segmentations. It is evident that the amplitude parameters a₁ and a₂ do not differ greatly across the normal, and liver disease models. The data in figure 6 also shows that, in general, segmentation had little effect on the relative values of the various model parameters derived.

Zoom In Zoom Out Reset image size

FAD model fitted parameters

Figure 7 shows the corresponding plots for the FAD analyses. It is evident that there is low statistical power in the cell segmentation data and, whilst the median values differ, they do so with considerable variability. These findings highlight the need for optimal imaging conditions in order to collect the most meaningful FAD data. Importantly, in the whole region and automated segmentation methods, where there is much more FAD data above the threshold photon counts per pixel than in cell segmentation, there is some evidence of two broad populations of bound FAD. Figure 7 also shows that FAD the τ₂ is longer in CCl₄ induced and Mdr2-/- fibrosis livers than in normal livers, with the FAD τ₂ ischemia-reperfusion at 6 h showing differences in the automated and whole region segmentation data.

Zoom In Zoom Out Reset image size

Metabolic ratios

Three commonly used metabolic ratios were calculated to assess changes in the liver: (1) redox ratio; (2) NAD(P)H amplitude ratio; (3) fluorescence lifetime redox ratio (FLIRR), that could further resolve differences between normal liver and disease state livers. The redox ratio of I_FAD/I_NAD(P)H showed a significant increase in both fibrosis models and HCC compared to normal livers irrespective of the segmentation method. Fatty liver disease was only significantly different from normal liver in the whole region segmentation and ischemia-reperfusion injury at 24 h was significantly increased for automated and whole region segmentation formats (figure 8). The NAD(P)H ratio of a₁/a_2NAD(P)H showed significant decreases for all disease models and all segmentation methods. The FLIRR of a_2NAD(P)H/a_1FAD showed significant decrease only in CCl₄ and HCC compared to normal liver for automated and whole region segmentation. With cell segmentation, there was no difference for all disease models compared to normal livers.

Zoom In Zoom Out Reset image size

Informed segmentation based on fluorescence lifetime

We used a sensitivity analysis of a τ_m,a based informed segmentation with a threshold photon count of 100 to quantify the spatial pixel distribution of collagen in the λ_ex 800 nm, λ_em 350 to 450 nm channel. Figure 9(a) shows the λ_ex 800 nm, λ_em 350 to 450 nm channel intensity images for our control liver and those for CCl₄, Mdr2-/-, fatty liver disease and HCC. Limiting τ_m,a to <2000 ps and photon count >100 resulted the loss of nuclei and cell boundaries (figure 9(b)) relative to figure 9(a). A further reduction in τ_m,a to <700 ps, with intensity >100 photons per pixel (figure 9(c)) led to the loss of almost all pixels from normal liver cells and the retention of pixels in regions that morphologically appeared to be collagen in the fibrosed liver (figure 9(a)).

Zoom In Zoom Out Reset image size

We then tested the hypothesis that these areas of fibrosis may be associated with cellular metabolic dysfunction by manual segmentation, in which we then arbitarily drew two areas of similar sizes corresponding to a fibrotic region (based on high collagen content) and a corresponding normal region (based on low collagen content), in the channel λ_ex 800 nm, λ_em 350 to 450 nm (figure 9(d)) and compared the NAD(P)H signals for the two regions from images at λ_ex 740 nm, λ_em 350 to 450 nm (figure 9(e)). We also included two regions of similar size from control livers for reference. NAD(P)H τ_m,a and NAD(P)H a₁/a_{2 NAD(P)H} ratio showed significant differences in the fibrotic region when compared to the control region (figures 9(f) and (g)). NAD(P)H τ_m,a increased and a₁/a_{2 NAD(P)H} ratio decreased in collagen rich areas in both CCl₄ and Mdr2-/- fibrosis models, whereas HCC showed the opposite results. Surprisingly, fatty liver disease did not show any difference in NAD(P)H τ_m,a and a₁/a_{2 NAD(P)H} ratio.

Technically challenging intravital MP-FLIM imaging has been performed on liver, skin, brain, and kidney to detect endogenous fluorophores and achieve analysis of metabolic cofactors NAD(P)H and FAD [9, 36, 45–50]. To date, most studies define NAD(P)H free as being in the cytoplasm and NAD(P)H bound being in the mitochondria and are associated with glycolysis and oxidative phosphorylation in the two sites, respectively. Remarkably, NADH and NADPH are involved in 488 human metabolic reactions which are not confined to the mitochondria for energy production, but are also located in the cytosol and nucleus in processes as wide ranging as transcriptional regulation and DNA repair [25]. Therefore, our understanding and interpretation of NAD(P)H and FAD MP-FLIM images are still developing. Importantly, alterations in the fluorescent lifetime of NAD(P)H free and FAD bound may not only reflect a shift in metabolic pathway but could also reflect either a change in lipid and protein synthesis and/or microenvironment changes, including variations in pH, concentration of oxygen, tyrosine, or tryptophan. Accordingly, limiting MP-FLIM analysis to the most commonly reported single metric of τ_m,a for NAD(P)H and optical redox ratio will provide only a limited understanding of how liver injury and liver disease has affected the redox state of the liver. Increasingly, both NAD(P)H free and bound metrics, including τ₁ and τ₂, along with multi-parametric analysis is likely to become more important [46, 51].

In order to examine the impact of the heterogeneity in the in vivo liver cells and spaces as a result of liver disease, this work has undertaken a more complete analysis of MP-FLIM data from normal and liver disease models than we have ever undertaken previously. In particular, this work has explored the impact of segmentation and disease state on the bi-exponential FLIM parameters for the endogenous fluorophores NAD(P)H and FAD, as well as in the λ_ex 800 nm, λ_em 350 to 450 nm channel, and the various metabolic ratios that can be derived from these parameters. The work extends our previous work in vivo which reported the most common metric for FLIM, the τ_m,a for NAD(P)H in hepatocytes [9]. The previous work showed the τ_m,a of NAD(P)H in normal liver (mean, stdev) was 1.511 ± 0.032 ns was estimated using λ_ex 740 nm λ_em 350 to 450 nm. Our FAD analysis was limited by our previous FLIM data being collected at λ_ex 800 nm and λ_em 515–650 nm [9]. Figure 1(d) shows that the more commonly used λ_ex 740 nm and λ_ex 890 nm avoid the small bleed through effect that may be seen from NAD(P)H in FAD estimation. The use of λ_ex 890 nm for FAD estimation may also be desirable in liver fibrosis to avoid overlapping fluorescence from elastin (2P λ_ex of 760 to 830 nm, an emission λ_em 475 to 575 nm and short and long lifetimes of 260 and 1960 ps [42]) that is often associated with liver fibrosis. Recently, Cao et al used a single λ_ex of 800 nm to imaging both NAD(P)H and FAD with emission filters of λ_em 450/50 nm and 560/80 nm, respectively, indicating this method may be preferred in vivo to decrease total imaging time, to avoid motion artifacts and increase temporal resolution, as needed in clinical settings [41].

Collagen emission is a second harmonic generated process with an effectively instantaneous fluorescence lifetime. With data collected at λ_ex 800 nm, λ_em 350 to 450 nm channel, we found a τ_m,a of about 1700 ps (figure 5(b)) reports of collagens (I, II and III) having τ_m,a of around 1.5 ns [52] or 0.29 ns and 1.68 ns at λ_ex 830 nm [53]. Visualized pixels with τ_m,a < 2000 ps (figure 9) did not have morphology consistent with that anticipated for collagen. The apparent loss of nuclear and cell membrane structures in images for 700 < τ_m,a < 2000 ps (figure 9) suggest that cellular components contribute to the apparent τ_m,a of ∼1700 ps shown in figure 5(b). We have shown elsewhere, using phasor analysis, that the τ_m,a for collagen in diseased livers is 87 ps [54] and the segmented images with τ_m,a < 700 ps appeared to have collagen like morphology. Further, it is evident that increased collagenization in fibrotic livers (figure 9) corresponds with an overall decrease in the τ_m,a for those same livers in the λ_ex 800 nm, λ_em 350 to 450 nm channel (figure 5(b)), consistent with the dominant effect of the essentially instantaneous collagen signal on the overall lifetimes.

In this report, we re-examined aspects of the data processing with different segmentation techniques and determined optimal parameter estimates of the fitted data to the heterogeneous datasets produced by different liver diseases. Multiple protocols used to obtain quality FLIM data [55–58], have mentioned the importance of calibration of the system, using fluorescein or rhodamine to estimate the correct IRF for model fitting and image analysis, and methods have been developed to estimate the correct IRF from reference images [59–61]. The most emphasised consideration of time-correlated single photon counting is the minimum amount of photons necessary for robust data fitting, which is generally 1000 photons per pixel and when not met, can lead to artefacts in the results [58]. We and others have been able to produce reliable model fitting to as low as 100 photon counts per pixel [62]. Many publications have shown that binning of data increases the confidence of model fitting without significantly change changing to the estimated parameters [35, 63, 64]. However, others have noted that, in heterogeneous samples such as diseased livers or the retina of the eye [65], binning can result in erroneous fluorescence lifetimes [66]. Lastly, the effect of segmentation was explored and we found that with fine segmentation of cytoplasm compared to nucleus compared to liver sinuses, we were able to detect differences in the NAD(P)H ratio of a₁/a_2NAD(P)H. However, when we compared segmentation of automated segmentation to a whole region centered in the image in FLIMFit, τ_m,a, τ_m,i, a_1, a_2, τ₁ and τ₂ did not vary significantly between segmentation methods for NAD(P)H and FAD. However, small differences could be discerned in τ_m,a and τ_m,i for the (figure 5(b)). The location of the NAD(P)H binding signal has always thought to be in the mitochondria, and that of free signal to come from the cytoplasm. However, one study using a higher single cell resolution, reduced binning and a pixel-wise two exponential fit has demonstrated that distinct NAD(P)H bound fluorescent lifetimes in adipocytes arise from the mitochondria as well as the cytoplasm [35]. This has led to the suggestion that the two short lifetimes seen in adipocytes represent folded (0.4 ns) and extended (0.9 ns) conformations of free NAD(P)H in solution whereas the two longer lifetimes represent two different binding states. It was further suggested that the ratio of free to bound NAD(P)H defines a glycolysis/oxidative phosphorylation ratio and a change in ratio reflects a change in enzyme binding [35]. Similarly, the evaluation of metabolic ratios using a higher resolution FLIM imaging of NAD(P)H, FAD and Trp found heterogeneity within the cells [67].

Importantly, Wallrabe et al [67], in measuring FLIM parameters for NAD(P)H, FAD and tryptophan in segmented, thresholded, xenograft cancer cell, reported that light-scattering and laser power needs in tissue sections made the FAD/NAD(P)H photon ratio (an intensity-based redox measurement) unsuitable for tissue sections. They suggested that the segmented, threshold, ROI approach using many data points to visualise differences between and within cells, as we have also described here, is preferred to the usual means based analyses. They also reported a heterogeneity and increase in FLIRR cellular responses after glucose stimulation. This simulates the aerobic glycolysis or Warburg effect seen in many tumor cells. However, high energy utilisation at the tumor edge of in vivo solid tumors may be accompanied by a hypoxic core, which may eventually become necrotic as a consequence. The stage and heterogeneity of the solid tumor will give rise to opposing FLIRR values dependent on the imaging site within the tumor.

We explored several combinations of NAD(P)H and FAD measured FLIM parameters based on theory developed in the field to capture the metabolic changes within cells in the liver during different disease states. In our analysis, we found that fibrosis, either in CCl₄ induced fibrotic hepatocytes or Mdr2-/- fibrotic biliary cells, could be distinguished by an increase in the redox ratio. Indeed, Mdr2-/- fibrosis is distinguishable through all 3 metabolic ratios assessed, irrespective of segmentation strategy used as there is an increase in the redox ratio, a decrease in the NAD(P)H ratio and an increase in FLIRR when compared to normal liver. These changes are due to an increase in τ₁ in both NAD(P)H and FAD indicating there are multiple metabolic pathways involved in this pathogenesis which modulate both NAD(P)H and FAD. In particular, a combined significant increase in redox ratio and decrease in τ_m,a for the the λ_ex 800 nm, λ_em 350 to 450 nm channel is a potential MP-FLIM predictor for HCC that may be useful in its diagnosis. In the ischemia-reperfusion injury model used here the bile duct was not clamped. In contrast, in other studies in which the portal vein, hepatic artery, and bile duct were clamped [2], there was an increase in FAD intensity and a decrease in NAD(P)H intensity [33] that could further distinguish a liver ischemia-reperfusion injury from other liver diseases.

Two other studies relevant to our analysis presented multi-parametric results to stratify metabolic response. In one case, an indicator of the metabolic responsiveness of the breast cancer to drug therapy over time and was based on an index that was the linear combination of three normalized parameters: redox ratio, NAD(P)H τ_m,a and FAD τ_m,a [68]. This approach tracks the metabolic changes over time and contrasts to the cross-sectional time data described here for different liver conditions. In another study, cells were subjected to hypoxia and glucose starvation, mitochondrial uncoupling, and fatty acid oxidation while 3 parameters: redox ratio, the NAD(P)H fluorescence lifetime, and mitochondrial clustering were used to distinguish distinct metabolic states in multiple cell types in vitro and in vivo [46]. This work parallels our goal to provide a combination of FLIM parameters that can provide a distinguishing signature for different liver injury and disease states for use as a diagnostic tool.

Recently, the validity of metabolic ratios reported in the literature from in vitro and in vivo studies has been questioned. In particular, the metabolic index NAD(P)+/NAD(P)H ratio has been challenged as a result of the assumed lifetime values for free and bound NAD(P)H, and differences in the NAD(P)H bound lifetimes reported from studies analyzing FLIM data with two, three, or four exponential fits versus the phasor approach [69]. Different FLIM acquisition hardware, acquisition considerations, as well as data analyses that can affect the reported NAD(P)H bound have been highlighted. For example, whether a full fluorescence decay is captured in the acquisition time window will affect the fitting, or whether the model fitting was applied globally or pixel-wise will ultimately affect the reported NAD(P)H bound lifetime. In line with this, we showed that variation in the data occur with the choice of programs to calculate the fit, when different IRF are applied and is dependent on the binning that is applied to the data. Efforts to validate the FLIM of NAD(P)H and FAD has been ongoing for many years. Studies have tried to separate NADH and NADPH using FLIM analysis to determine the relative contribution [29]. Indirect measurements using mass spectroscopy has shown that the expression of NAD and/or NADP differs between various tissues [70–72]. A normalized optical redox ratio FAD/(NAD(P)H + FAD) has also been used to measure the cell redox changes associated with stem cell differentiation and the optical redox ratio was correlated to the gold-standard liquid chromatography/tandem mass spectrometry (LC/MS) measurements of intracellular cofactor concentrations. Surprisingly, the optical redox ratio FAD/(NAD(P)H + FAD) correlated to the LC/MS measurements of NAD(P)⁺/(NAD(P)H + NAD(P)⁺) but did not correlate to the LC/MS measurements of FAD/(NAD(P)H + FAD) [73]. The optical redox ratio has also been shown to correlate well to the rate of oxygen consumption measurements made on the metabolic analyzer (Seahorse XFe24, Seahorse Bioscience) when ETC disruptors were applied [74]. An in vivo FLIM study of the brain systematically applied metabolic inhibitors to the brain surface under a cranial window to discern differences in NAD(P)H during metabolic perturbation to glycolysis, the TCA cycle, the ETC, or oxidative phosphorylation [36]. A careful analysis of multiple NAD(P)H FLIM parameters was used to detect differences between the aforementioned metabolic states.

We have shown that MP-FLIM is a sensitive and useful method to interrogate changes in metabolism and that segmentation based on other structural or functional markers such as collagen could aid in identifying distinct cellular subpopulations. This work is relevant to recent single cell RNAseq studies showing liver cell populations can be stratified into many functional subsets [75, 76]. This work has also used Becker & Hickl TCSPC data recording on a Dermainspect MPM in our in vivo real time clinical imaging approach. Others have proposed a wide-field, time-gated FLIM approach based on only four images, with high precision being achieved by either a summing of frames or by use of pixel-binning [77]. However, our work here has shown that spatial resolution may be compromised by pixel binning, leading to erroneous results in heterogeneous samples and the summing of frames may also need to be corrected for intravital motion [10].

The authors have declared that no conflicting interests exist. This work was supported by grants from the National Health and Medical Research Council of Australia (Grants: 1049906; 1049979, 1002611, 1126091, 1107356). Many thanks to Leslie Burke for a careful review of the manuscript. The authors from the Diamantina Institute are based at the Translational Research Institute (TRI), Queensland, Australia.

We thank Dr Svetlana Rodimova, a researcher at the Privolzhsky Research Medical University, Russia, who inspired much of this paper through her request for us to expand on our 2015 Real-time Histology in Liver Disease paper [9] by publishing the relative contributions of a₁ and a₂ for NAD(P)H and the fluorescence life-time data for FAD, for comparison with their work [33, 34], and who also reviewed this paper. We also implicitly acknowledge Dr Sean Warren, the author of FLIMFit software tool developed at Imperial College London that was used for the analysis of most of the images described here. We also express our appreciation to Dr Wolfgang Becker, the developer of the TCSPC technology used in the FLIM analysis and of SPCImage5.5 that was used to collect the TCSPC data for individual images. Lastly, we are grateful to Drs Hauke Studier and Wolfgang Becker of Becker & Hickl Pty Ltd for providing us with an evaluation copy of SPCImage 8 for use in this work.