1. Introduction

Molecular imaging techniques have been widely used for assessing biochemical and biological processes at the molecular level in live intact subjects during preclinical and clinical targeted drug development [1]. Nuclear (PET and SPECT) and optical imaging modalities are the two main technologies employed to monitor targeted drug delivery efficacy and pharmacokinetics, non-invasively and longitudinally [2]. Nuclear imaging techniques are highly sensitive and can provide whole body 3D imaging of a drug-ligand labeled with a radiotracer. Though, they offer only millimeter-scale spatial resolutions, require costly equipment, and are limited to monitor only one biomarker at a time [3]. Conversely, fluorescence optical imaging offers the potential to deploy affordable imaging platforms with small form factors, simultaneously track multiple biomarkers of interest via probe multiplexing, and leverage a large panel of available probes while presenting similar sensitivity to nuclear imaging platforms. Hence, optical imaging has made remarkable strides in supporting the targeted drug development pipeline [4]. However, when applied to in vivo imaging, optical imaging performances are limited by scattering – leading to decreased resolution during imaging of deep tissues. Thus, monitoring in vivo parameters that are critical to the assessment of drug delivery efficacy, such as drug-target engagement and intra-tumoral heterogeneity (ITH), remains extremely challenging [5].

Over the past two decades, Fluorescence Lifetime Imaging (FLI) has attracted increasing attention as it provides a unique ability into monitoring the tissue microenvironment (e.g., pH [6], viscosity [7], etc.) and metabolic status [8]. Moreover, FLI methodologies can uniquely provide means to monitor protein conformation and interactions when leveraging Förster Resonance Energy Transfer (FRET) [9]. Of particular interest herein, this so-called FLI-FRET approach enables monitoring of drug-target engagement in live intact subjects [10]. FLI is compatible with traditional fluorescence imaging approaches and can be integrated with well-established platforms, including widefield, confocal and multi-photon systems [5,11]. Though, such implementations are typically limited to studies of cell cultures or require implantation of intravital chambers for interrogating shallow tissues in live animals (a few hundred microns depth) [12]. Imaging 3D intact biological samples at greater depth and over large field of views (FOVs) requires taking into account the scattering nature of the photons collected. This can be performed via Fluorescence Molecular Tomography (FMT) [13].

Over the last decade, FLI capabilities have been implemented in FMT for improved sensitivity/specificity [14] as well as to monitor FRET interactions [15,16]. Still current implementations are performed for macroscopic investigation of tissues and provide only millimeter scale resolution [17] – hence, are not well suited to resolve ITH. Recently, a new implementation of diffuse optical tomography, named Mesoscopic Fluorescence Molecular Tomography (MFMT) has been proposed to image biotissues at millimeters depth (mesoscopic regime) and with higher resolution (hundreds of microns) than FMT. This new methodology has found applications in diverse biomedical fields, including bioprinted scaffolds [18], glioblastoma model [19], mouse cortical activity [20], and tumor xenografts [21–23]. Thus, MFMT is highly promising for reporting on the in vivo ITH in 3D. Though, to date MFMT has only been implemented in continuous-wave (CW) mode with intensity-based acquisition [24,25]. Still, MFMT is poised to extend its utility with time-domain (TD) augmentation for target engagement assessment based on lifetime quantification.

Herein, we report on the instrumental development and characterization of a time-domain MFMT (TD-MFMT) system that enables simultaneous evaluation of fluorescence lifetime contrast, as well as volumetric reconstruction of the distribution of deep-seated fluorescent molecular probes, within turbid media. Specifically, our TD-MFMT system is equipped with a pulsed laser exciting from visible (VIS) to near-infrared (NIR) range and two detections: 1) an sCMOS camera for CW acquisition with high spatial density; and 2) 16-channel SPAD array coupled with a time-correlated single-photon counting (TCSPC) module for TD acquisition. We report on the key performances of the system along its various instrumental dimensionalities, including spatial, spectral and temporal characteristics, as well as application-focus features, including sensitivity and specificity. This extensive experimental characterization and validation is performed using controlled fluorescent well-plate samples. Furthermore, we demonstrate the platform’s utility to perform multiplexed 3D imaging in a tissue-like phantom across a large spectral range. Altogether, we show that TD-MFMT is capable of providing unique 3D distribution of fluorophores compared to macroscopic FLI (MFLI) [26] as well as enhanced spatial resolution compared to functional diffused optical tomography (FDOT) [27]. These characterization experiments provide the foundation for further studies focusing on investigating in vivo drug-target engagement and ITH longitudinally.

2. System design and characterization

2.1 Optical setup

The system scheme is depicted in Fig. 1 a). It implements two types of pulse lasers as the illumination sources to deliver wide-spectrum VIS-NIR light or high-power NIR light, respectively. The type of laser illumination is selected based on the imaging scenario. For in vitro imaging (i.e. bioprinting applications), we employ a supercontinuum laser source (SuperK Extreme EXR-20, NKT Photonics) to produce pulses at 78 MHz repetition rate in the wavelength range of 480-1200 nm. A multi-wavelength filter (SuperK SELECT, NKT Photonics) based on acousto-optic tunable filter (AOTF) technology is applied to allow wavelength-tunable light for multiplexed imaging. Despite relatively low output power (< 5 mW/nm), this laser is capable of exciting multiple endogenous and exogenous biomarkers across VIS to NIR range. For in vivo and ex vivo imaging, the illumination is switched to a tunable Ti:Sapphire pulse laser (MaiTai, Spectral Physics) to deliver high-power NIR light. This laser source can generate pulses from 690 nm to 1040 nm at 80 MHz rates and illuminate samples at power close to (but below) the Maximum Permissible Exposure (instructed by the ANSI laser safety standards).

 

Fig. 1. System scheme and spatial characterization results. a) TD-MFMT system setup. b) Full image of USAF target. c) Intensity profile of the blue dash line. Red dash square: distinguishable finest line pairs (49.6 µm spacing). Scale bar: 1 mm. d) Paradigm for reflective imaging geometry. e) Paradigm for tomographic imaging geometry. f) Representative TPSFs captured at 700 nm in the scheme of time-resolved diffused spectroscopy. g) Normalized measured multi-mode and simulated intensities of the diffusive detectors.

Download Full Size | PDF

After being delivered via fibers and followed by collimation, the laser beam is directed through the first linear polarizer (P1) (LPVISC100, Thorlabs) and reflected by a polarizing beamsplitter (pBS) (48-545, Edmund Optics). Next, the s-polarized beam is raster-scanned by a galvanometer scanner (GM) (GVSM002/M, Thorlabs) and subsequently sent into a scan lens (SL) (LSM03-VIS, Thorlabs). The angle of the s-polarized beam changes as the GM steers, producing a step-by-step scanning pattern on the image plane. The reflected and back-scattered signal returns through the same path as the incident light. Another crossed-polarized linear polarizer (P2) is used to minimize the specular reflection of the object surface. Then, the collected signal goes through a filter wheel (FW) (CFW6, Thorlabs) for specific spectral separation. The emission signal is split into two detection paths by a 50:50 beam splitter (BS) (BSW10R, Thorlabs) and focused on two individual detectors via two achromatic lenses (L1 and L2, AC254-080-AB, Thorlabs).

In the detection paths, an sCMOS camera (Zyla 4.2, Andor Technology, UK) is set up to measure fluorescence intensities with high spatial sampling, while a customized SPAD detector (Politecnico di Milano, Italy) coupled with TCSPC module concurrently records the time-resolved data sets. The sCMOS camera features 2048$\times$2048 pixel resolution with 6.5 µm pixel size and high quantum efficiency (QE) (82$\%$ at maximum) over VIS-NIR range. In contrast, the SPAD consists of a 32-channel linear array, in which half (16 channels) are connected to the TCSPC, set at a 1.6 ps time step resolution and photon counts quantized at 32-bit. Each of the 16 detector channels is 50 µm in diameter and the pitch size is 250 µm. Restricted by sensor design and manufacturing at present, SPAD technology is inferior to the established CCD and sCMOS cameras with regards to fill factor and QE, resulting in lower photon collection efficiency [28].

Our TD-MFMT system can work either in reflective imaging geometry or tomographic imaging geometry, depending on whether the target application calls for collecting only the on-axis light (epi-configuration) or including the back-scattered signal with different source-detector separations (see the paradigms in Fig. 1(d) and (e)). Overall, the spatial data set concurrently acquired is CW: 2048$\times$2048 pixels and TD: 16 pixels per scanning point. Herein, we typically use a 200 µm scanning step over a 6$\times$6 mm$^2$ area, leading to 900 scanning points. Hence, we can collect spatially dense data sets, though in practice, the CW data sets is reduced to improve SNR (by binning) but also due to computational limitations regarding the size of the inverse problem that needs to be solved. For reflective imaging, the on-axis super pixel is binned by $8\times 8$ original pixels of the sCMOS camera and a single channel (typically Channel 1) of the SPAD is aligned with the epi-configuration. The size of the super pixel is 52 µm, which, for fair comparison, is comparable to that of one SPAD detector channel. For tomographic imaging, the back-scattered photons are collected by multiple super pixels of the camera as well as the off-axis channels of the SPAD (Channel 2-16) with millimeter-range separations from the laser source.

2.2 Spatial parameters characterization

Since spatial resolution is key for scanning-based imaging techniques, the lateral resolution of TD-MFMT was characterized by raster-scanning an USAF 1951 resolution test target (R3L3S1N, Thorlabs) in reflective geometry. Figure 1(b) shows the full image of the resolution target. We set the scanning position as 80$\times$80 and the step as 50 µm, corresponding to the total number of pixels and each pixel scale, respectively. As the reflective signal at each position is obtained from a single super pixel, the spatial resolution is dependent on the scanning step. The blurring of the fine line pairs is mainly attributed to the finite size of the illumination spot. The line pairs in the red rectangle (Element 3 in Group 3), with line width and spacing of 49.6 µm, are distinguished in Fig. 1(c), which proves that the spatial resolution is better than 50 µm at least. Note that despite the larger line spacing of the Elements 2 and 1 (55.7 and 62.5 µm respectively), each line occupies space with more than one scanning step, leading to a relatively less distinguishable result. Although the spatial resolution could be potentially improved by configuring smaller steps, we have not explored the limitation herein as we are moreso focused on the characteristics associated with sensitive fluorescence lifetime quantification.

For optical tomography, the MFMT configuration leads to very large variations in the intensities to be acquired over the full FOV. To emphasize the large variations in intensity, we provided the profile of signal intensities at the excitation wavelengths that could be expected from scattering tissues and as computed using a Monte Carlo forward model [29]. Such high variations are far greater than what typically encountered in FMT due to the nature of the photons collected in the mesoscopic regime, both minimally- and multi-scattered ones. In turn, this creates difficulty in implementing FLI with gold standard TD imager, i.e, gated ICCD cameras, as they are limited to 12-bit data depth due to the amplifier component. Conversely, SPAD arrays are well known to provide increased dynamical range. To demonstrate the potential of our SPAD array to capture adequate signals over the whole FOV, we experimentally acquired excitation data over a scattering phantom with the exact same optical properties as the ones used in the MC simulations. For the time-resolved data, the temporal point spread functions (TPSFs) (Fig. 1) were time-integrated to obtain the intensities. The reduction of the intensity in Fig. 1 demonstrates the decreasing tendency of photon collection as the source-detector (s-d) separation increases. Since the data depth of sCMOS is 12-bit (maximal intensity is 4096 and slight saturation at the maximum intensity) and typical background noise is ∼100 (with 100 ms exposure time), the intensity variation of sCMOS is theoretically inferior to the one of SPAD (32-bit).

Of note herein, we used a voxel-based Monte Carlo approach to compute the forward model and Jacobian used for the following 3D reconstructions. The boundary conditions were set to semi-infinite planar. The use of a voxel-based model enables to leverage the symmetry of the descanning approach to greatly diminish the computational burden. By selecting voxel sizes that are multiplicative of the scanning steps, we can limit ourselves to simulate the forward model for one source and all detectors as described in [30]. Such boundary conditions can be maintained in vivo by using a non-invasive imaging window such as proposed in [31]. To ensure high accuracy in the forward model, we calibrated the physical s-d separations in the back-scattered measurements based on the dimensions of the pixel size (sCMOS) and the pitch size (SPAD), as well as the system’s magnification factor, to ensure the registration between the forward model and the back-scattered measurements.

2.3 Time resolved characteristics

The main temporal characteristic of a time-domain imaging system is typically reported by the full width at half maximum (FWHM) of the instrument response function (IRF) [32–34]. Since IRF quantifies the system response between the laser pulses as an input and the collected photons by detectors as an output, its features depend on the temporal characteristics of the pulsed laser source and the time-resolved detector. This IRF temporal characterization is conducted with two laser sources (SuperK and MaiTai) independently to account for different imaging scenarios.

Herein, the system IRFs were measured similarly to our group’s prior work [35,36] – placing a sheet of diffusing paper on the imaging plane and individually recording the reflections for each detector channel at different wavelengths. However, depending on the end user’s target application, one may consider using Spectralon to ensure no autofluorescence is collected during IRF acquisition [37]. To protect the SPAD from saturation and the pile-up effect, this measurement was carried out with low laser power (< 0.2 mW/cm$^2$) by switching the filter wheel to a neutral density filter (NE20A-B, Thorlabs) to attenuate the laser output. During the characterization of each channel, the specific channel was aligned at the on-axis position (epi-configuration) to collect the reflective photons from the paper. Representative IRFs acquired using two different laser sources are plotted in Fig. 2(a) and (c) across six wavelengths. The FWHMs are calculated by interpolating the corresponding IRFs with 1.6 ps resolution to quantify the temporal resolution of the system (shown in Fig. 2 (b) and (d)). The FWHMs can achieve ∼90 ps using SuperK laser (except for 650 nm at the edge of AOTF available wavelength range) and ∼120 ps by MaiTai throughout their respective spectral ranges. In addition, Fig. 2(d) shows the FHWM results across the channel array with illumination at 700 nm, demonstrating comparable temporal resolution of all channels.

 

Fig. 2. Temporal characteristics of the system. a) IRFs recorded by Channel 1 using the two lasers in NIR range. b) FHWMs of the IRF captured by Channel 1 across the available spectral range. c) IRFs recorded at 700 nm by all channels using the two lasers. d) FHWMs of the IRF captured at 700 nm by all channels.

Download Full Size | PDF

2.4 Sensitivity characterization

Fluorescence imaging modalities enable highly sensitive detection of fluorescent probes at the molecular level. To validate the sensitivity and specificity of the TD-MFMT system, we captured multi-mode (CW and TD) data sets in reflective geometry via illumination with the SuperK. Afterwards, we compared the intensities of the two modes and resolved the lifetimes of exogenous and endogenous fluorescence dye solutions in controlled multi-well settings.

Alexa Fluor 700 dye (AF700, A20010, Thermo Fisher Scientific) was diluted in phosphate-buffered saline (PBS) solution at different concentrations: 10, 30, 100 and 300 nM. Using 700 nm laser excitation, CW and TD measurements were sequentially acquired by the two detectors with the emission filter (FF01-740/13-25, Semrock), under the control of a specialized LabVIEW program. A CW intensity value and a TPSF were recorded at each position by the two detectors respectively. All TPSFs were processed by pixel-wise iterative least-squares fitting using Alligator [34,38]. Each TPSF, corresponding to a single scanning point, was integrated for SPAD-based intensity value, and fitted into a mono-exponential decay convolved with the IRF for retrieval of the corresponding lifetime result. Two groups of AF700 well-plate samples were imaged based on the aforementioned acquisition and processing workflow, investigating the influence of limited photon collection, as well as decreased fluorophore concentration, on the accuracy of lifetime quantification. In Group 1, we imaged 300 nM AF700 solutions with different dwelling time (200, 400, and 600 ms) for obtaining a varied degree of photon collection. In Group 2, three AF700 solutions with gradient concentrations (100, 30, and 10 nM) were imaged with comparable collected photon counts acquired by adjusting the dwell time. The CW-mode and SPAD-based fluorescence intensity images recorded with the two detectors, the fitted lifetime maps, as well as the representative TPSFs and the corresponding fitting results are shown in Fig. 3(a) and (d). The resulting lifetime distributions (Fig. 3(e)) demonstrate that the lifetime of AF700 is resolved accurately and in agreement with our previous research ($\tau \approx 1.0 ns$) [38], even in the cases of very low photon count (∼17,000 counts/second for the whole array) (Group 1, 200 ms) and nanomolar fluorophore concentration (Group 2, 10 nM). Of note, a small increase in lifetime is seen as the concentration decreases. Beyond environmental changes associated with the sequential imaging of the well, we hypothesize that the increase in noise exhibited in the TPSF as concentrations decrease, will bias slightly the full curve fitting towards higher lifetime. Still, this change is minimal – as the mean lifetime changes less than 60 ps across the three wells. Additionally, as integration times per dwell time is increased, laser and electronic jitters can lead to slight temporal instabilities and less accurate timing with the employed IRF.

 

Fig. 3. Sensitivity and specificity characterization results with well-plate settings. Multi-mode intensity and lifetime/FD maps of a) AF700 series with different acquisition parameters (pixel = 400, scanning step = 250 µm); b) FRET quantification using AF700-AF750 FRET pair with different A:D ratios (pixel = 900, scanning step = 250 µm); c) Autofluorophores (FAD, Rb and PPIX) (pixel = 400, scanning step = 250 µm). d), h) and i) are representative normalized IRFs, fluorescence decays and mono-exponential fits for AF700, FAD and PPIX. e) Corresponding fitted lifetime distributions for AF700 series in box plot. f) Representative normalized decays for the different AD ratios. g) Corresponding $f_{FRET}$ distributions. Scale bar: 1 mm

Download Full Size | PDF

To evaluate the system’s sensitivity, we utilized the F-value (also referred to as photon-economy) as the metric of merit. The F-value is defined as $F=\sqrt {N}{\sigma }/{\tau }$, where ${\sigma }/{\tau }$ is the ratio of standard deviation to the mean value of the repeated lifetime measurements, and N represents the number of detected photons. The more dominant the Poisson noise is within the measured fluorescence decay, the closer F-values are equal to 1 – representing how efficient the system is with the collected photons [39–41]. As illustrated in Fig. 3(a), the F-values in both groups (e.g. 1.91 in Group 2 at 100 nM) indicate the high sensitivity of the system. Of note, the poor F-value (Group 2, 10 nM) could be caused by systematic perturbation (such as laser instability and shot noise accumulation) due to the long dwell time required to obtain sufficient photon counts.

Furthermore, the sensitivity characterization was extended to the quantification of FRET donor fraction (i.e. FD or $f_{FRET}$), which enables nanoscale drug-target engagement imaging assays and examines the system’s capabilities for imaging in complex environments containing multiple lifetime species. $f_{FRET}$ is defined as the donor fraction engaging in FRET out of the total donor molecules including the non-FRETing portion. One notable characteristic of FRET is a reduction in the measured donor fluorescence lifetime when in close proximity to an acceptor molecule (typically within 10 nm [42]). Hence, $f_{FRET}$ is typically quantified via bi-exponential fitting [43]. A well-plate sample was prepared by the AF700-AF750 FRET pair in PBS buffer, with AF700 Mouse IgG (MG129, Thermo Fisher Scientific) as the donor dye and AF750 goat antimouse IgG (A-21037, Thermo Fisher Scientific) as the acceptor dye. Four wells were filled with AF750-AF700 FRET dye combinations at four acceptor-to-donor (A:D) concentration ratios ranging from 0:1 to 3:1. The multi-mode data sets were collected using the same excitation and the emission filters as for the above AF700 experiment. However, since FLI-FRET data are comprised of two exponential terms (the quenched and unquenched donor contributions), the TD data were processed using a bi-exponential fitting model in Alligator. Initial lifetimes of donor-only ($\tau _1$ = 1.0 $\pm$ 0.05 ns) and quenched donor ($\tau _2$ = 0.35 $\pm$ 0.05 ns) were applied for the fitting based on prior research [44]. Similarly, SPAD-based intensity images, the fitted FD maps, and representative TPSFs for different A:D ratios are shown in Fig. 3(b) and (f). As expected, we observe an increasing trend of the $f_{FRET}$ with increasing A:D ratio (Fig. 3 (g)), which further supports the system’s ability to monitor drug-target engagement.

2.5 Specificity characterization

A unique advantage of optical imaging, and especially FLI, is the ability to image multiple fluorophores or fluorophore states in the same sample [45]. This advantages provide unique approaches to monitor biological tissues with high specificity. The TD-MFMT system characterized herein can leverage both an extended spectral range for efficient color-based multiplexing while using FLI for unmixing fluorophore emitting in the same spectral range (in our case FRET donor, see above). Herein, we validated the ability of our TD-MFMT system to image different fluorophore species with large differences in lifetimes by acquiring data from three autofluorescent samples, including Flavin Adenine Dinucleotide (FAD, SIGMA F6625, Midland Scientific), Riboflavin (B2) (Rb, 47861, Sigma-Aldrich), and Protoporphyrin IX (PPIX, P562-9, Frontier Scientific). These three fluorophores were diluted by different solvents and prepared at different concentrations. FAD and Rb were dissolved by distilled water (FAD: 150 µM, Rb: 20 µM), while PPIX was diluted by a mixture of dimethylformamide (DMF, 227056, Sigma-Aldrich) and methanol with the volume ratio of 1:1 (PPIX: 10 µM). Given that the fluorescence brightness is intrinsically linear with respect to the product of the quantum yield and the extinction coefficient, the rationale for modulating the fluorescence concentration was to obtain a comparable level of emission intensity (and correspondingly, a similar SNR) for fair quantitative comparison across all three fluorophores. The illumination wavelength was selected by the AOTF unit at 490 nm, considering the absorption efficiency of the fluorophores and the laser output power distribution in the spectral sense. An excitation filter (FF01-488/50-25, Semrock) was placed after the illumination light to eliminate potential spectral bleed-through in the emission signal. The multi-mode measurements were acquired with a longpass emission filter (FF01-515/LP-25, Semrock) following the same imaging protocol as that used for the sensitivity characterization experiments previously. The averaged lifetimes of the three autofluorophores measured herein are 2.32 $\pm$ 0.12 ns, 2.76 $\pm$ 0.21 ns and 7.83 $\pm$ 1.51 ns for FAD, Ribloflavin and PPIX, respectively. These results are within the broad range of autofluorescence lifetime reported in [46]. The distinguishable contrast between the lifetime maps (Fig. 3 c)) confirms the system’s specificity for simultaneously monitoring and discriminating multiple fluorophores with similar emission spectra based on lifetime contrast.

3. Diffuse phantom study for 3D imaging validation

To test the system performance for fluorescence tomographic imaging, we performed a phantom study mimicking a fluorescent region embedded into in vivo tissue environment. Agar powder (05040-250G, Sigma-Aldrich) was used to construct the solid phantom and Intraplid (20 $\%$ emulsion, Sigma-Aldrich) was introduced to control the scattering coefficient. No absorption agent was involved into the medium for this validation experiment. As shown in Fig. 4(a), two paralleled capillaries ($D_{outer}$ = 1 mm, $D_{inner}$ = 0.75 mm) filled with two autofluorescence dyes (10 µM PPIX, 20 µM Riboflavin), were positioned at the same depth in an agar phantom (optical properties: $\mu _a$ = 0.002 mm$^{-1}$, $\mu _s$’ = 1 mm$^{-1}$). A third capillary containing 5 µM AF700 was placed beneath in an orthogonal orientation.

 

Fig. 4. Phantom study validation of multi-mode reconstruction using phantom. a) Spatial diagram of three capillaries filled with fluorescence dyes in an agar phantom. b) SPAD-based intensity and the corresponding resolved lifetime maps for two acquisitions captured by Channel 2 (Scale bar: 1 mm). c) 3-D surface of the three capillaries in different colors based on threshold segmentation of the reconstructed fluorescent CW data sets (yellow: Riboflavin, purple: PPIX, blue: AF700). d) Localization of the capillary glass walls in the phantom using the Micro-CT scanner. e) Fused 3D image of c) and d). f) and g) show the left-side view and front view of e).

Download Full Size | PDF

Multi-mode measurements were carried out at two illumination wavelengths (490 nm and 700 nm) consecutively, acquiring the excitation and emission data with frame rates of 5 Hz (200 ms/frame) over a 6.6 $\times$ 6.6 mm$^2$ area. The principle of CW-mode data collection and reconstruction follows as laid out in our previous work [47]. Briefly, we employed the Monte Carlo (MC) simulator [29] to create a forward model and formulate a sensitivity matrix for photon propagation in scattering tissues. The 3D distribution of the effective fluorescence quantum yield (x) was retrieved by solving the inverse problem of Ax = b, where A represents the sensitivity matrix and b denotes the back-scattered measurements. The sCMOS data at each scanning steps were processed to yield super-pixels. Following our prior work, we down-sampled the sCMOS data to 49 super-pixels, leading to a total of 44,100 source detector pairs in the measurement space (very high density) [48]. A classic iterative approach, conjugate gradient (CG), was applied as the inverse solver in the reconstruction. In this study, the voxel size of the reconstructed volume is 200 µm. Moreover, the absorptive geometrical information of the phantom was obtained with a Micro-CT scanner (MARS Bioimaging, New Zealand) featuring with 90 µm spatial resolution for the cross-validation. We utilized Imaris 9.9 software (Oxford Instruments, UK) for 3D visualizing the isosurfaces generated by threshold segmentation for the MFMT (CW-mode) and Micro-CT reconstruction images. As for the TD data sets, a measured temporal decay in time-resolved fluorescence tomography basically consists of early- and late-arriving photon portions. Since the timescale for photon scattering in biological tissue is much smaller than the one for fluorescence lifetime, the late-arriving portion decay can be factorized into a product of spatial sensitivity weights and pure exponential decays. Thus, fluorescence lifetime can be directly resolved by exponentially fitting the measurements [49]. Herein, we mono-exponentially fit the tail portion of the TPSFs acquired by Channel 2 of SPAD with a source-detector separation of 125 µm.

The reconstruction result of three capillaries by CW data sets presents accuracy on depth and geometry as validated by Micro-CT (Fig. 4(c) - (g)). The deeper capillary was reconstructed evenly at depth of 2 mm, which demonstrates the system’s capability of performing 3D imaging on vascular-like fluorescence objects in thick biotissues at millimeter(s) depth with high fidelity. Figure 4(b) depicts the normalized SPAD-based intensities and the corresponding mono-exponentially fitted lifetime map for two acquisitions. Two autofluorescence capillaries are comparable in intensity and uncertain to distinguish. Additionally, more scattering effect is observed in the intensity map of the deeper AF700 capillary. Overall, the combination of 3D imaging based on high-density CW data and FLI contrast enables accurate retrieval of fluorophore distribution while also distinguishing them. These results pave the way for implementation of the system for volumetric assessment of ITH and drug-target engagement simultaneously in vivo.

4. Ex vivo imaging

Having validated the system with controlled well-plate and phantom studies, we extended the performance evaluation to ex vivo assessment to highlight its potential for investigating outstanding biological challenges. A nude mouse carrying HER2-positive xenograft model of human ovarian cancer (SK-OV-3) was intravenously injected with AF750 and AF700 conjugated with Trastuzumab (TZM) – a clinical anti-HER2 monoclonal antibody (40 µg AF750-TZM and 20 µg AF700-TZM, A:D = 2:1). After in vivo imaging using a gated-ICCD apparatus (details elsewhere in Ref. [36]), the mouse was euthanized 48-hour post-injection, and the tumor xenograft was extracted for ex vivo imaging.

Multi-mode imaging was performed in tomographic geometry with the same excitation wavelength and the emission setting as the sensitivity characterization section used. A high-resolution reflective image was additionally shot by the sCMOS camera with widefield setup (Fig. 5(a)). We observed the distinct spatial distributions in a single SPAD-based intensity map (Fig. 5(b)) as well as among the maps acquired by different channels with different s-d separations (Fig. 5(d)-(f)). This distinction exhibits the depth-dependent heterogeneity of tumor morphology and demonstrates the system’s potential to resolve the imaging of ITH in a depth-dependent manner. For FRET quantification, TPSF decays (Channel 1 at epi-position) were fitted bi-exponentially using Alligator. The donor-only and quenched donor lifetimes refer to our previous work ($\tau _1$ = 1.3 ns, $\tau _2$ = 0.5 ns) [34]. The $f_{FRET}$ quantification was benchmarked against a well-established time-resolved imaging platform, MFLI system [34]. The FRET fraction results obtained by both systems are in high agreement as shown in Fig. 5(h) (mean $\pm$ standard deviation: 2.2 $\pm$ 1.5$\%$ (TD-MFMT) versus 2.8 $\pm$ 4.2$\%$ (MFLI)).

 

Fig. 5. Ex vivo Trastuzumab-HER2 tumor xenograft imaging. The xenograft sample was imaged in tomographic geometry (pixel = 625, scanning step = 400 µm, exposure time = 500 ms). a) Widefield reflective image. b) SPAD-based intensity map acquired by Channel 1 at epi-position. c) Fitted $f_{FRET}$ map. d) - f) SPAD-based intensity maps acquired by different channels at different s-d separations (Channel 5: s-d = 0.5 mm, Channel 9: s-d = 1 mm, Channel 13: s-d = 1.5 mm). g) Representative IRF and the fitted decay. h) $f_{FRET}$ distributions quantified by TD-MFMT and benchmarked by MFLI system. Scale bar = 2 mm.

Download Full Size | PDF

5. Discussion

Preclinical imaging assessment, and especially imaging of small animals bearing tumor xenografts remains a critical step in the drug development pipeline. However, current modalities are not well suited to characterize key drug delivery parameters in vivo that are increasingly recognized as crucial: ITH [50] and target engagement [51]. Herein, we report on a new imaging platform designed to reveal both ITH in vivo via high-resolution 3D optical tomography in intact tissues and target engagement via FRET donor lifetime monitoring. Through the combination of a sCMOS and SPAD array, both high-density CW data and time-resolved data can be acquired concurrently for seamless registration. As highlighted by the experimental results, the spatio-temporal and spectral characteristics enable the system to sense the nanomolar-level fluorophores as well as quantify fluorescence lifetime both in the case of mono-exponential and challenging bi-exponential cases (FRET fraction) at high spatial and temporal resolution. Moreover, the preliminary results of the ex vivo tumor xenograft imaging support the system’s potential to resolve tumor heterogeneities in depth-dependent manner within relevant biological tissues. Of importance, even if SK-OV-3 is a HER2 positive cell line, it has been reported to show very small level of target engagement with TZM [52,53]. We have also reported similar results recently [34]. SK-OV-3 tumors displayed reduced TZM uptake both by MFLI and Immunohistochemistry (IHC). They appeared more resistant to drug delivery, in agreement with their disrupted vasculature and collagen-rich tumor microenvironment. Our FRET fraction quantification herein support these prior findings. It also underlines that even if drug accumulation is still exhibited within a tumor 48-hrs after injection, this observed distribution alone cannot be automatically associated with targeted cellular drug delivery. This is an important finding, as drug efficacy is directed linked to successful target engagement. We will carry out additional in vivo studies with immunochemistry validation to further establish the unique potential of our platform. Still, our current approach suffers from a few limitations.

First, the instrument design and imaging protocols herein have room for improvement. In tissues, the current acquisition time is mainly limited by the low fill factor of the SPAD array and overall collection efficiency. We are currently using a scanning lens optimized for the visible range. We expect to improve collection efficiency by replacing it with a NIR one. Also, our current SPAD array incarnation does not benefit from a microlens array, which would significantly increase its performance. Moreover, the acquisition of TD data sets can be further improved by identifying the lowest photon counts rates acceptable for accurate lifetime quantification. This can be further optimized by temporally binning the data as gating improve SNRs and decrease acquisition times while not affecting NIR-FRET quantification accuracy as described in [33].

Second, our system is capable of acquiring dense CW data sets. Though, herein we did not fully leverage this very high sampling rate for optical tomography. As MFMT is a computational imaging approach that relies on solving a scattering inverse problem, some considerations and limitations can arise while using such very large data sets. Chiefly, the Jacobian matrix to be manipulated can become extremely large such that the CG implementation used herein becomes impractical on a personal computer (out of memory). Hence, following previous work, we have down-sampled the 2048$\times$2048 CW data set to only 49 super pixels. Similarly, the image space was discretized to 200 µm – limiting the attainable resolution. We will continue to further investigate alternative approaches to solve the inverse problem for improved resolution by increasing the CW data set employed and refining the image space discretization. These alternative approaches could include algebraic methods that are memory efficient [54] and computationally competitive thanks to their GPU implementation [55], informed data reduction [56] and/or direct Deep Learning based reconstructions [57,58]. Similarly, we will investigate the potential improvements in including the temporal data sets in the inverse problem as previously demonstrated [17,59].

6. Conclusion

This work reports on the characterization of a TD-MFMT platform that is geared towards in vivo imaging of tumor xenografts. Overall, the experimental results demonstrate its ability to work over large spectral range and lifetime values for imaging and characterizing small concentrations of fluorophores. Moreover, a preliminary ex vivo study exhibits its potential to reveal ITH in tissues as well as reporting on target engagement. Hence, this novel imaging platform is expected to be a valuable tool for monitoring targeted drug delivery and associated biological questions in preclinical oncology.