Improved in Vivo Whole-Animal Detection Limits of Green Fluorescent Protein–Expressing Tumor Lines by Spectral Fluorescence Imaging

Abstract

Green fluorescent protein (GFP) has been used for cell tracking and imaging gene expression in superficial or surgically exposed structures. However, in vivo murine imaging is often limited by several factors, including scatter and attenuation with depth and overlapping autofluorescence. The autofluorescence signals have spectral profiles that are markedly different from the GFP emission spectral profile. The use of spectral imaging allows separation and quantitation of these contributions to the total fluorescence signal seen in vivo by weighting known pure component profiles. Separation of relative GFP and autofluorescence signals is not readily possible using epifluorescent continuous-wave single excitation and emission bandpass imaging (EFI). To evaluate detection thresholds using these two methods, nude mice were subcutaneously injected with a series of GFP-expressing cells. For EFI, optimized excitation and emission bandpass filters were used. Owing to the ability to separate autofluorescence contributions from the emission signal using spectral imaging compared with the mixed contributions of GFP and autofluorescence in the emission signal recorded by the EFI system, we achieved a 300-fold improvement in the cellular detection limit. The detection limit was 3 × 10³ cells for spectral imaging versus 1 × 10⁶ cells for EFI. Despite contributions to image stacks from autofluorescence, a 100-fold dynamic range of cell number in the same image was readily visualized. Finally, spectral imaging was able to separate signal interference of red fluorescent protein from GFP images and vice versa. These findings demonstrate the utility of the approach in detecting low levels of multiple fluorescent markers for whole-animal in vivo applications.

M METHODS FOR CELL TRACKING, imaging gene expression, or monitoring tumor growth in mice exist for essentially all small animal imaging modalities. Each technique has advantages, and the particular method chosen in individual experiments depends on the question at hand. ¹ The vast number of molecularly targeted fluorescent agents ² and the broad range of fluorescent proteins available for imaging, ³ combined with the ability to image multiple targets simultaneously, ⁴ are distinct benefits that enhance the useful information that can be derived from fluorescent optical images.

Optical imaging devices with a single excitation wavelength band exciting and single emission band recording whole-animal fluorescence from the same side are the most straightforward devices conceptually and have been used robustly in hundreds of applications.^5–9 A limitation of such devices is that naturally occurring fluorescence from tissues (autofluorescence) is more pronounced in the visible spectrum compared with near-infrared optical imaging and can confound the recorded signal. Although the spectral shape of signal intensity versus wavelength is quite different for autofluorescence compared with fluorescent proteins such as green fluorescent protein (GFP) when taken across the entire emission spectrum, a single band emission recording cannot separate these components if a single excitation wavelength is used and techniques such as lifetime imaging ¹⁰ are not employed. This limits the amount of GFP that is detectible owing to autofluorescence contributions to the recorded signal. One approach to overcoming this overlap of emission spectra is to acquire multiple images at different wavelengths to create emission spectra on a pixel-by-pixel basis and to assign contributions to various predetermined component spectra, such as GFP and autofluorescence. This may be performed by minimizing the residual spectrum after removal of the appropriately weighted components using a least squares fitting approach. ¹¹

The goals of this study were to determine the number of subcutaneous GFP-expressing cells that could be visualized using a spectral imaging approach from a commercially available system (Maestro, CRi, Woburn, MA) compared with the detection threshold using an optimized home-built system that employed single excitation and emission filters selected to minimize bleedthrough and other sources of unwanted signal. Although detection depends on various factors, including levels of GFP expression between different transfected cell lines, we were readily able to image 3 × 10³ subcutaneously implanted cells using the spectral imaging approach compared with a threshold of 1 × 10⁶ equally dispersed identical cells using a single bandpass emission technique. We additionally evaluated the dynamic range of the number of cells that could be seen and the ability to visualize GFP-expressing cells mixed with red fluorescent protein (RFP)-expressing cells and conversely the ability to visualize RFP-expressing cells mixed with GFP-expressing cells.

Materials and Methods

Spectral Curve Acquisition

The transmission properties of the bandpass filters used in spectral imaging and standard epifluorescence imaging were measured on a Varian Cary 50 Bio UV-VIS spectrophotometer (Palo Alto, CA) from 400 to 800 nm. The excitation and emission spectra for enhanced GFP were acquired with the use of a fluorometer (Hitachi U-4500, Tokyo, Japan). The emission spectrum was acquired at an excitation wavelength of 434 nm at a range between 470 and 650 nm. The excitation spectrum at a range between 400 and 530 nm was acquired using a fixed emission wavelength. The skin autofluorescence spectrum was acquired on a Yvon-Horiba Florolog Model 3 fluorometer (Edison, NJ) at an excitation wavelength of 425 nm. An emission spectrum was acquired by scanning between 430 and 900 nm. The transmission properties of the liquid crystal tunable filter were acquired by a Nicolet FT spectrometer (Thermo Electron Corporation, Waltham, MA).

Cell Culture

Mouse CT26 colon carcinoma and rat 9L gliosarcoma cells were obtained from the American Type Culture Collection (Manassas, VA). The CT26-GFP tumor cell line was generated by stably transfecting CT26 cells with a recombinant retrovirus encoding enhanced GFP (eGFP), and the 9L cell line was stably transfected with either eGFP or RFP. CT26 cells were cultured in RPMI 1640 supplemented with 10% fetal bovine serum (FBS), l-glutamine (2 mM), HEPES (10 mM) (Cambrex, Walkerville, MD), sodium pyruvate (1 mM), glucose (14 mM), and penicillin (100 U/mL)/streptomycin (100 μg/mL). 9L cells were grown in Dulbecco's Modified Eagle's Medium, 10% FBS, l-glutamine (2 mM), and penicillin (100 U/mL)/streptomycin (100 μg/mL). The cell lines were cultured at 37°C in a humidified 5% CO₂ atmosphere. Unless otherwise noted, all materials were obtained from Cellgro (Herndon, VA).

For cell implantation, tumor cells in culture were detached using HyQ trypsin (Logan, UT) and centrifuged. The cells were resuspended in medium and counted. Cell population was determined by counting non–trypan blue–stained cells on a hemocytometer. Fluorescent cells were analyzed via flow cytometry to determine the mean fluorescence intensity of the cell population. Dilutions of different cell numbers, from 1 × 10³ to 3 × 10⁶, were prepared depending on the experiment. Cell populations that were not 100% fluorescent were accordingly compensated.

Animal Models

Female nude mice (nu/nu, Cox-7, Massachusetts General Hospital, Boston, MA) were anesthetized with isoflurane/O₂. Mice were housed and maintained under aseptic conditions according to guidelines set by the Institutional Animal Care and Use Committee.

Detection Limit CT26-GFP by Spectral Fluorescence Imaging

A series of CT26-GFP cells, ranging from 1 × 10³ to 3 × 10⁶ cells, were injected subcutaneously onto the dorsum of female nude mice. Fluorescence imaging was conducted immediately following injection of the cell line. Imaging was conducted using a whole-animal imaging system (Maestro). Multiple images were acquired with fixed bandpass excitation and emission filters in place. The liquid crystal tunable filter of the system was placed in series after the bandpass emission filter, which rejected the vast majority of reflected excitation light and decreased the out-of-band rejection required by the tunable filter. The center peak of the liquid crystal tunable filter was incremented between images from 500 to 800 nm centerline in steps of 5 nm for GFP imaging. Each image was acquired with identical exposure time. To prevent saturation of pixels in any image of the set, the tunable filter was set to the emission peak of GFP (530 nm). This wavelength resulted in the strongest signal, and the exposure time was set to have the maximum pixel intensity at this wavelength at 70% of the charge-coupled device (CCD) saturation. Each series of images was subsequently analyzed as described below.

Epifluorescent Continuous-Wave Single Excitation and Emission Bandpass Imaging

Female nude mice were subcutaneously injected with a series of CT26-GFP cells on the dorsum, ranging from 1 × 10³ to 3 × 10⁶ cells. Fluorescence imaging was performed immediately postinjection on a whole-animal epifluorescent continuous-wave single excitation and emission bandpass imaging (EFI) system developed in-house and optimized for GFP excitation and emission wavelengths, similar to previously published systems. ⁴ The system included a 12bit Pixelfly QE CCD (PCO, Germany) mounted over a ring light for a uniform excitation field and housed in a light-tight box. Excitation light from an external 300 W xenon lamp was filtered through a bandpass filter (450–480 nm; Omega Optical, Brattleboro, VT) before entering the ring light. The emission spectrum was also filtered through a bandpass (500–530 nm, Omega Optical) in front of the camera. The exposure time was 0.3 seconds for GFP imaging, and both bandpass filters were displaced for white light images. Quantitative measurements were performed on the acquired epifluorescent images using a circular region of interest (ROI) placed over each focus of cells and reported using image analysis software (OsiriX, Apple, Geneva, Switzerland).

Dynamic Range Determination

To establish if dynamic range, as determined by the ability to detect a small focus of GFP-expressing cells in the same animal as a larger number of GFP-expressing cells in a second focus in the same image, was limited using spectral imaging, two cell populations were subcutaneously injected onto the dorsum: 1 × 10⁴ and 1 × 10⁶ CT-26 GFP cells. Animals were imaged on the spectral imaging system. The exposure time was set such that the 1 × 10⁶ injection was 70% of CCD saturation at 530 nm, and image sets were taken between 500 and 800 nm at 5 nm steps. Spectral separation was performed, as below, with experiments done in triplicate.

GFP/RFP Spectral Separation

To evaluate the ability to separate GFP and RFP signals, a mixture of 9L-GFP and 9L-RFP cells was injected subcutaneously. For signal detection of 9L-GFP, a series of 9L-GFP cells, ranging from 1 × 10⁴ to 3 × 10⁶ cells, was mixed with 1 × 10⁵ 9L-RFP cells. As with the CT26-GFP detection limit experiment, the tunable filter was set to 530 nm, and the exposure time was set to maximally fill the CCD to 70% of saturation. Image sets were acquired between 500 and 800 nm in 5 nm steps and were used to spectrally separate GFP and RFP signals.

For signal detection of 9L-RFP, a series of 9L-RFP cells, ranging from 1 × 10⁴ to 3 × 10⁶ cells, was mixed with 1 × 10⁵ 9L-GFP cells. The tunable filter was set to 583 nm center frequency, and the exposure time was again set to 70% CCD saturation. Image sets were taken between 550 and 800 nm center frequency at each 5 nm step and analyzed as below to evaluate signal separation.

Spectral Analysis

Spectral analysis was performed in several steps. Initially, regions of skin without GFP- or RFP-transfected cells were defined as autofluorescence and the relative signal intensity of these regions recorded by the CCD at each center frequency wavelength was determined. Additional curves of signal intensity (SI) versus center frequency wavelength were then obtained from regions containing GFP (+ autofluorescence of overlying skin) and RFP (+ autofluorescence of overlying skin). The obtained autofluorescence curve was subtracted from the other two curves to generate curves of signal intensity versus wavelength, which represented GFP or RFP contributions, respectively, to the signal. Image analysis software, coupled with the spectral imaging system (Maestro), was used to generate these curves, including computed GFP and RFP curves. The signal intensity curves across the range of center frequencies were subsequently divided for each pixel into contributions from these spectral components, and generated images were computed with SI reflecting the contributions of the respective different components. A least squares optimization was performed to separate the component signals. ¹¹

Images corresponding to GFP or RFP signals were used to quantify fluorescence signals. Measurements were performed on the computed images using a circular ROI placed over each GFP, RFP, or GFP-RFP mixture injection (representing signal) and several noninjected areas (representing background autofluorescence signal). For both single bandpass EFI and spectral imaging, background signal (B) for a particular ROI was calculated as the signal multiplied by the ROI area (A_ROI) divided by the exposure time. Two types of background signal were calculated and used for analysis: (1) the average background signal was calculated as the average of these signals (B_avg) and (2) the average background signal for a ROI was calculated as (1) and multiplied by the ROI area (A_ROI). Method 1 was used to determine the signal for a fluorescent region, whereas method 2 was used to report the background signal of the animal relative to the ROI area. Signal (S) for a fluorescent region was reported in counts per second, whereby the signal was divided by the exposure time. The total fluorescent signal was then calculated by the following equation:

${\mathrm{S}}_{\text{GFP},\text{RFP},\text{GFP}/\text{RFP}}=({\mathrm{S}}_{\text{GFP},\text{RFP},\text{GFP}/\text{RFP}}-{\mathrm{B}}_{\text{avg}})\times {\mathrm{A}}_{\text{ROI}}$

Results

GFP Filtering

The filter spectra for the EFI system (Figure 1) show an excitation bandpass filter and an emission bandpass filter that fit within the spectral excitation and emission peaks for GFP, illustrating an optimized single bandpass emission system used for imaging. The two filters are separated such that bleedthrough of excitation light is minimized while much of the true GFP signal is captured. The filter spectra for the spectral imaging system also show an excitation bandpass filter that fits within the excitation peak of GFP. For the emission filter, however, a longpass filter, instead of a bandpass filter, was used. A main objective was to remove as much of the excitation light as possible so that required out-of-band rejection was minimized for the tunable filter. Out-of-band rejection for the tunable filter is approximately 10⁻². The longpass emission filter transmits light from 515 nm and longer wavelengths. The tunable filter has a transmission window of 10 to 20 nm. Spacing of 5 nm center frequency between images results in overlapping contributions to images from adjacent wavelengths. Note that using a single bandpass emission filter, separation of autofluorescence from true GFP signal is not possible; thus, signal intensity in the recorded image reflects a combination of both contributions.

CT26-GFP Detection Limits of Spectral Imaging and EFI

Female nude mice were subcutaneously injected with a series of CT26-GFP tumor cells and imaged on the spectral imaging and EFI systems, as described above. Nine injections of each cell population were used for spectral imaging, and three injections for each cell population were used for EFI. As seen in Figure 2, spectral imaging resulted in the CT26-GFP signal increasing linearly across the range of cell populations with an R² factor of .98. The lowest detectable cell population, 3 × 10³ cells, was more than 10-fold higher than that of background for spectral imaging. For EFI, the lowest detectable cell population above background signal was 1 × 10⁶ cells. All lower cell populations for EFI were within the standard error of the mean of background autofluorescence. Similar signal intensities across this range of cell number, combined with Figure 1, suggest that autofluorescence dominates the recorded signal in this range so that the weaker GFP signal is not seen.

Figure 1.

A, Green fluorescent protein (GFP) excitation (blue) and emission (red) with excitation (orange) and emission (green) filter sets for a single bandpass emission system. Skin autofluorescence emission curve (purple) relative to emission curve of GFP is also shown. Filter sets were optimized to include as much of the excitation and emission curves as possible while maintaining optimal separation to minimize filter crosstalk. B, GFP excitation (blue) and emission (red) with excitation (orange) and emission (green) filter sets for the spectral imaging system. Skin autofluorescence emission curve (purple) relative to emission curve of GFP is also shown. Excitation and emission filters were also optimized for spectral imaging. In contrast to A, the spectral imaging system uses a longpass emission filter rather than a bandpass filter. This allows the rejection of the majority of excitation light but allows discrete spectral acquisition throughout the emission range. The longpass filter was placed in series with the liquid crystal tunable filter. The center frequency of the tunable filter was incremented in steps of 5 nm.

Figure 2.

Fluorescence intensity versus cell number for single bandpass epifluorescence and spectral imaging systems. The skin autofluorescence values (background signal determined from areas devoid of green fluorescent protein–labeled cells) for each system are shown on the graph for comparison. For epifluorescence imaging (EFI), the detection threshold is approximately 1 × 10⁶ cells owing to contributions to signal from skin autofluorescence. For spectral imaging, values are linear across the range of different cell numbers (R² = .9863) and at 3 × 10³ cells, significantly higher (p = .01) than the background signal.

Additional Spectral Components Can Further Reduce Misassigned Contributions to GFP Signal

Initial spectral assignment of GFP and autofluorescent signals in mice with subcutaneous injections of CT26-GFP often leads to adequate separation and appropriately increasing recorded focal GFP signal with increasing numbers of GFP cells. However, as seen in Figure 3, some mice exhibited additional focal skin contaminants that required additional spectral separation to appropriately isolate the GFP signal. This situation became more important in cases in which fewer GFP-positive cells were present. Using only two spectral curves to represent GFP and autofluorescence, the background signal in some places was as much as 1.4 times higher than that of the 1 × 10⁴ GFP cell injection (see Figure 3). For these mice, adding a third spectral curve specific to these contaminants lowered the misassigned contribution to the GFP spectrum image, resulting in a lower detection limit.

Dynamic Range

To determine if weak GFP signals could be seen in the same computed GFP images as stronger signals, despite contributions to raw spectral images from autofluorescence, mice were injected subcutaneously with 1 × 10⁴ and 1 × 10⁶ GFP-expressing CT26 cells and imaged from 500 to 800 nm. As seen in Figure 4, the computed images had sufficient dynamic range, even with the required separation of pixel values into the GFP and non-GFP spectral contributions, to visualize the 100× dynamic range of cell amounts, without any saturation of the computed image at the more concentrated GFP focus.

GFP-RFP Mixtures

Detection limits for the number of GFP-expressing cells depend in part on the relative GFP expression between cell lines. Owing to lower GFP expression, the detection limit for 9L-GFP was higher than CT26-GFP. The lowest number of detec Table 9 L-GFP cells was 1 × 10⁴ cells with a signal 1.9 times higher than the background signal (Figure 5A). Owing to somewhat overlapping spectra, detection of GFP in the setting of RFP expression is expected to decrease, raising the minimum numbers of cells required for visualization. In the presence of 1 × 10⁵ 9L-RFP cells, the detection limit increased to 1 × 10⁵ 9L-GFP cells. However, in this case, the SI was 35 times higher than the background signal (Figure 5B). A subcutaneous injection of 3 × 10⁴ 9L-GFP cells (mixed with 1 × 10⁵ 9L-RFP cells) was 1.7 times higher than background (data not shown), but the standard error of the mean was within noise limits.

The detection limit for 9L-RFP was similar to the 9L-GFP cells. The limit was 1 × 10⁴ 9L-RFP cells for the lowest number of detectable cells with a signal 4.1 times higher than background (Figure 5C). In the presence of 1 × 10⁵ 9L-GFP cells, the detection limit was 1 × 10⁵ RFP cells and 4.3 times higher than background signal (Figure 5D). A subcutaneous injection containing 3 × 10⁴ cells was 1.5 times higher than background (data not shown) but was within the noise limits of the system.

Discussion

Murine fluorescent protein imaging has been approached using a variety of techniques, from whole-animal imaging to confocal microscopic evaluation of tissues of interest. In this article, we showed that sampling across different emission wavelengths and computing relative contributions of fluorochromes to resultant images markedly improved the detection threshold of GFP-expressing cells compared with single bandpass emission techniques for whole-animal evaluation. The detection of 3 × 10³ GFP-expressing cells for subcutaneous detection approaches limits proposed for subcutaneous bioluminescence detection of a few hundred luciferase-expressing cells. ¹² In addition, these results are comparable to GFP cell detection by intravital microscopy ¹³ and external imaging of cancerous lesions. ¹⁴ Fluorescence imaging provides additional benefits, including ease of multiplexing fluorophores and direct high-resolution correlation of findings with fluorescent microscopy. Moreover, higher expression of GFP per cell or a more concentrated focus of cells would further improve the achieved in vivo detection threshold shown here; smaller injection volumes were not used to help ensure accurate delivery of the expected number of cells. Two of the cell lines used had different levels of GFP expression: the CT26-GFP cell line had a higher expression level (2,162 arbitrary units) than the 9L-GFP cell line (965 arbitrary units). This difference of fluorescence intensity led to a lower detection limit of CT26-GFP than 9L-GFP (3 × 10³ CT26-GFP cells versus 1 × 10⁴ 9L-GFP cells). In our laboratory, we have observed alternate generated cell lines that have higher expression levels of GFP than these lines and thus would likely have a lower detection threshold. Alternatively, this method would also be useful for detection of somewhat higher numbers of low-GFP-expressing lines, which may not have been previously detectible using EFI or other systems. Applied microscopically, this method would also improve the visualization of such low-GFP-expressing cell lines during histologic analysis.

Figure 3.

A, Spectral separation of green fluorescent protein (GFP) and autofluorescence using two spectral curves (GFP and skin autofluorescence) may result in a residual signal that is misassigned as GFP, altering the levels recorded, owing to focal skin contamination noted in some mice. The addition of a third spectral curve representing these contaminants was added for spectral separation and yielded more accurate GFP signals (B).

Figure 4.

A, White light image of mice injected with two different cell populations: 1 × 10⁴ and 1 × 10⁶. Black fiducial markers are present to note placement of injection. Note that there is no injection at the middle black fiducial marker. B, Fluorescence intensity for the two injected green fluorescent protein populations. C, Merged image of the fluorescence and white light images. Relatively low cell counts (10⁴ cells) were appropriately visualized in the setting of 100-fold higher cell counts in the same image, despite autofluorescent signal. There was no saturation of the image at the site of higher counts in either raw or computed images.

Figure 5.

A, Detection limit of the 9L-GFP cell line. This cell line exhibits lower mean fluorescence intensity than the CT26-GFP cell line, with a resultant higher detection limit of 1 × 10⁴ cells. B, In the presence of 1 × 10⁵ 9L-RFP, the detection limit is 1 × 10⁵ 9L-GFP cells. C, Detection limit of the 9L-RFP cell line. This cell line exhibits the same detection limit as 9L-GFP of 1 × 10⁴ cells. D, In the presence of 1 × 10⁵ 9L-GFP cells, the detection limit increases to 1 × 10⁵ 9L-RFP cells. GFP = green fluorescent protein; RFP = red fluorescent protein.

Autofluorescence has many sources: chlorophyll in food and collagen and elastin in skin. By modifying the diet for animal subjects to contain chlorophyll-free food, autofluorescence owing to diet can be minimized ¹⁵ ; however, skin autofluorescence would remain. With spectral imaging, each component of autofluorescence can be separated and removed from the desired signal.

As with luciferase imaging, the depth of cells from the surface increases attenuation of signal using the method demonstrated here. The primary advantage gained using the spectral imaging approach is the removal of confounding autofluorescence signal from true GFP contributions. Other methods, such as tomographic techniques, ¹⁶ are being introduced for deeper imaging and have approached the problem differently. In those studies, the animals are immersed in a matching fluid, and the authors have reported detection of 50,000 cells in a small volume. In addition to the liquid crystal tunable filter, other approaches for generating frequency-selective images, such as multiposition filter wheels, gratings and prisms, and laser-scanning single point spectrographs, should produce results similar to those obtained here.

The spectral imaging technique also lends itself to separating fluorescent proteins or fluorochromes, which partially overlap, as shown with mixed cell populations of GFP- and RFP-expressing cells. Appropriate separation of spectra into components depends in part on similarities in the shape of the emission curves for the components. More dissimilar regions improve robust assignment. Fortunately for GFP imaging, the autofluorescence profile shown in Figure 1 differs considerably from the GFP emission. Owing to this markedly different profile, the removal of autofluorescence contributions has allowed an approximately 300-fold improvement compared with reflectance techniques in which autofluorescence remains. Such improved detection thresholds may allow improved visualization of early tumor growth and imaging of immune response modulation to superficially implanted tumors.