Distinguished photons: increased contrast with multispectral in vivo fluorescence imaging

Introduction

Over the past decades, imaging has become a critical component of medicine with striking advances in magnetic resonance (MR)-and computed tomography (CT)-based imaging methods. For the most part, however, these imaging modalities reveal anatomical rather than molecular features, and while this has proved extremely useful, many of the molecular and cellular changes that occur at the onset of a disease are not detectable with purely anatomical imaging. Even when disease states become evident in anatomical imaging methods, the information can be hard to interpret, since it is not directly related to molecular states such as protein or gene expression. The recent focus on and development of techniques for molecular imaging, wherein markers that can be imaged with specially design instruments and that specifically highlight cells and tissues with particular molecular characteristics, aim to provide this critical tool for research and clinical application.

In preclinical phases of the drug development cycle and during much tumor biology and antibody development research, small-animal imaging plays a pivotal role. The ability to perform relevant, minimally invasive or noninvasive imaging helps reduce costs and enables longitudinal studies of multiple processes and parameters in individual animals. Several imaging modalities, most notably optical imaging, positron emission tomography (PET), and magnetic resonance imaging (MRI), have been adapted to facilitate these studies through the use of specialized contrast agents and imaging probes that either identify the existence of certain genetically modified cells or tissues or that produce signals roughly proportional to regional molecular abundance or the rate of specific events in molecular pathways (1–8). In human subjects, optical imaging is limited to relatively superficial targets by the absorbing and scattering properties of tissue in the visible and near-infrared (NIR). Consequently, PET and MRI are arguably the preferred modalities for ultimate clinical application. However, in small-animal models, such as mice and rats, the shorter path-lengths allow sufficient photon flux to reach the surface of the animal (again, especially in the NIR), enabling fluorescent and bioluminescent emissions to be detected with acceptable signal-to-noise ratios and using relatively low-cost instrumentation compared with PET- or MRI-based systems. MRI is also limited both by poor sensitivity and by the limited diversity of contrast agents. PET is more sensitive, but while there is a wide range of available radiotracers, the use of PET in the laboratory is limited by the difficulty and cost of producing these short-lived reagents. In contrast, probes and reporter proteins based on fluorescence and bioluminescence are already widely used in molecular biology laboratories, do not present ionization hazards, and are physically and chemically far easier to work with in the laboratory (2).

Optical Techniques: Bioluminescence Versus Fluorescence

Optical molecular imaging systems and methodologies have been developed that use both bioluminescent (9) and fluorescent (10) signals. Bioluminescent systems typically use the firefly luciferase gene as a reporter. The major attraction of this approach is that although absolute light levels generated by the targets may be low, signal is produced only where luciferase is present, leading to extremely low background signals. In contrast, fluorescence-based imaging requires an external light source to stimulate the emission of light from the probe and may be accompanied by bright background signals arising from intrinsic autofluorescent species.

Choosing an Appropriate Detection System

The instrumental challenges of imaging bioluminescent and fluorescent signals from small animals are quite different. Bioluminescence is typically regarded as being a zero-background technique (although there are very small amounts of natural bioluminescence present) with an extremely low photon flux from the reporter gene (11). This requires that the imaging system incorporate an extremely light-tight enclosure, a very sensitive camera capable of long exposure times (occasionally as long as 20 min), and collection optics with a low f-number to collect as many photons as possible.

The limiting factors in a bioluminescence imaging system are typically background light leaking into the enclosure, camera read-noise, and camera thermal noise. Background light leakage can be essentially eliminated through the careful engineering of a light-tight enclosure. Camera read-noise and thermal noise can be substantially reduced or even eliminated by using a high-end charge-coupled device (CCD) camera with a high quantum efficiency (QE) and a very low (down to −105°C) operating temperature. Unfortunately, these CCD cameras are extremely expensive and may require liquid nitrogen to cool, adding complexity and cost. In addition, low f-number optics to collect the most photons translates into a very shallow depth-of-focus for the camera, which, for a three-dimensional object such as a mouse, can lead to focus issues.

An additional challenging aspect of bioluminescence-based methods is that the intensity of emission is dependent on the availability of substrate in the environment surrounding the luciferase gene-expressing cells. This luminescent substrate is administered typically by introducing it, and various routes can affect biodistribution and clearance kinetics. The timing of image acquisition with respect to the introduction of the substrate can thus affect reproducibility.

Despite these limitations, well-engineered bioluminescence imaging systems have been on the market for a number of years, and the technical problems with the methodology and its sensitivity have been mainly solved.

The major limiting factor for in vivo fluorescence imaging, by contrast, is fluorescence from the animal itself. Such tissue autofluorescence is a ubiquitous signal that is present regardless of the excitation wavelength used. Autofluorescence is strongest when excited in the blue range of the spectrum and weaker when excited in the red or NIR (700–800 nm) range, but is the usually the limiting factor for sensitivity and limits of detection for all wavelengths. Autofluorescence in tissue is primarily from components in skin (collagen, which fluoresces in the green) and food [chlorophyll breakdown products, which fluoresce in the red (12)], although muscle, organs, and tumors all also autofluorescent, mostly in the visible, but also out into the NIR regime (13).

Typically, tissue autofluorescence is bright enough to fill a significant portion of the dynamic range of a standard scientific-grade CCD camera in seconds, thus making thermal noise, which is the accumulation of thermally generated electrons on the CCD element over time, insignificant even for near-ambient CCD cameras. This leads to a key advantage of fluorescence in vivo optical imaging techniques—aggressively cooled, expensive CCDs are not required. Fluorescent signals, which here can refer either to a desired signal, such as a green fluorescent protein (GFP), or to autofluorescence, can be bright, even to the point of being visible in the presence of normal room light. Thus, in many cases, camera exposure times for fluorescence-based imaging are limited to a few seconds at most, because the CCD pixel wells will quickly saturate. This is in contrast to the situation with bioluminescence, in which arbitrarily long exposure times (sometimes many minutes) can be required to achieve sufficient signal. The effect of cooling (aggressive versus moderate) on sensor-related signal-to-noise is illustrated in Table 1, which looks at the effect of profound cooling (−105°C) in a short and a long exposure regime when the desired signal generates 1000 electrons at the sensor in either case.

Table 1.  Effect of Cooling on CCD-Based Signal-to-Noise in Long- and Short-Exposure Regimes with Signal Strength Set to 1000 Electrons

The variables compared here include varieties of noise that degrade system precision. “Noise,” as used here, denotes the statistical uncertainty (scatter) contributed by a process, not its effect on mean intensity. For example, dark current is a time- and temperature-dependent increase in CCD signals that is unrelated to the presence of light. While the mean value of the increase can be subtracted from the overall signal, uncertainty about this mean value means introduces fluctuations in repeated measurements under identical conditions. The same is true for shot-noise (based on sampling uncertainty related to the light-based signal at the CCD) and camera read-noise, which is the noise contributed by the electronic process of reading out the CCD. Read-noise is independent of exposure time and signal intensity and is modestly affected by the temperature of the sensor and related electronics. Table 1 shows that the cooling of a sensor affects overall sensor-related noise contributions only when exposure times are long (minutes) and has virtually no effect when they are short (seconds), the situation that prevails in most fluorescence measurements. In the case illustrated, signal-to-noise doubles from 15:1 to 30:1 when the camera is cooled to -105°C in the 5-min case, whereas the ratio remains at about 30:1 regardless of cooling level in the 5-s case. Values used for these calculations are derived from specifications of cameras based on the Kodak KAF-1062E sensor (Eastman Kodak Company, Rochester, NY, USA).

Another major source of measurement imprecision is shot-noise arising from background signals. Shot-noise (also called the Poisson counting noise) equals the square root of the number of electrons counted and is a fundamental statistical limitation common to every imaging sensor. Filling the well of typical 12-bit scientific-grade CCD cameras with well-depths of some 14,000 electrons will generate shot-noise of approximately 120 electrons. This again implies that aggressive cooling of a CCD camera to reduce read-noise and thermal noise (which even for long exposures amount only to some 50 to 100 electrons) will have little benefit. Thus, if one is trying to measure a fluorescence marker signal in the presence of autofluorescence, the shot-noise of the total signal limits the precision with which the marker emission can be quantified, especially when the marker emission is significantly less than autofluorescence. For example, if the total number of electrons (reflecting photons arising from signal and autofluorescence) collected by a CCD pixel is 10,000, the shot-noise will be 100 electrons. If the marker fluorescence represents only a few hundred of these events, total shot-noise will be of the same order as the magnitude of the signal.

One means by which the signal-to-noise performance in this regime can be improved is to overcome the well-depth limit of the CCD by taking multiple images of the same scene and adding these images together digitally. For signals that are small relative to the background autofluorescence, this will improve the signal-to-noise ratio by approximately the square-root of the multiplication factor. The drawback of this approach is the longer measurement times involved. However, by imaging fluorescence-based samples using data collection times usually employed for bioluminescence (moving from 5 s to 5 min, for example), the signal-to-noise will be enhanced by approximately an order of magnitude, improving the detectability of faint signals correspondingly.

Increasing Contrast with Multispectral Imaging

There are two primary methodologies for removing the interference from tissue autofluorescence in in vivo imaging. The first involves utilizing fluorescence lifetime information—tissue has a different fluorescence lifetime than the fluorophores used in biomedicine, and on this basis, the signals can be separated. However, the pulsed lasers typically used for this type of imaging can be expensive. The second means of removing interference from tissue autofluorescence is by using multispectral imaging (MSI) and unmixing to separate and isolate the contributions from each of the fluorophores in a sample.

Figure 1 shows two spectra taken from the skin of a mouse that has been injected with an antibody labeled with a quantum-dot fluorophore. The majority of the skin shows a spectrum (green spectrum) that compares well to the spectrum of autofluorescence taken from a negative control mouse. A spectrum taken from over the subcutaneous tumor in the left leg of the mouse (red spectrum) shows an increase in fluorescence at 630 nm, which is the maximum emission wavelength of the quantum dots used in this example (14). It is clear from comparing the spectra that even if an extremely narrow range bandpass filter is used (shaded gray region in graph), the majority of the photons collected at 630 nm will come from the autofluorescence of the mouse rather than from the fluorophore of interest.

MSI has been shown to be an effective means of separating the contributions to the total fluorescence of each of the fluorophores in a sample (15,16). The unmixing algorithms used with MSI systems allow the generation of images of the isolated fluorescence from the fluorophores in the sample. Figure 2 shows a comparison between the monochrome image of the mouse at 630±15 nm (left) and the result of unmixing the tissue autofluorescence (not shown) from the contributions from the quantum dot (right). Because the quantum dot fluorescence has been separated from the autofluorescence, the quantum dot appears as a bright white signal on a black background, showing the greatly increased signal-to-noise of the unmixed image.

Spectral unmixing benefits measurements of label fluorescence in two ways. First, there is the obvious advantage of improved measurement accuracy through the appropriate partitioning of the optical signal into its various sources [target(s) and autofluorescence(s)], and second, the presence of shot-noise will affect the overall accuracy of spectral unmixing by adding uncertainty to the mathematical matrix-inversion procedure (17). However, exposures are taken at multiple wavelengths, thereby effectively increasing the number of electrons detected, and reducing the effective shot-noise. The degree to which this is a benefit depends on how distinct (orthogonal) the spectral vectors are that reflect the various fluorescent sources in a sample.

Materials and methods

All data were collected using a Maestro™ In-Vivo Imaging System [Cambridge Research & Instrumentation, Inc. (CRI), Woburn, MA, USA]. The mice were anesthetized, using either ketamine [200 mg/kg injection, intraperitoneal (i.p.) single bolus, approximately 5 min before imaging] or halothane gas, and placed inside the imaging system. Nude mice are convenient to image because of their lack of hair. On the other mice, the hair can be removed by shaving or preferably with a depilatory cream (e.g., Nair™).

MSI data sets (cubes) were acquired with images typically spaced every 10 nm throughout the desired spectral range. The mouse shown in Figure 1 and 2 was imaged from 580 to 700 nm using a green excitation filter (503–555 nm) and a 580 nm long-pass emission filter. The mouse imaged for Figure 3 and 4 was imaged using the same filter combination and scanning from 580 to 720 nm. The mouse imaged for Figure 5 was imaged twice using two difference excitation/emission filter pairs, and the data from those two collections were concatenated into one larger data set. The first filter set was a blue excitation (445–490 nm) with a 515 nm long-pass filter, and the second filter set was the aforementioned green filter pair. In general, depending on the Maestro model, fluorescence emission ranges from 420 to 720 or from 500 to 950 nm are achievable, with filter pairs being available to cover the entire range.

RGB (red-green-blue) color images were synthesized from the spectral cube by mapping the spectral data into those color channels. Either true color (in which spectral regions are mapped faithfully into their corresponding RGB channels) or false color displays can be generated; the latter are useful when signals in the NIR (by definition invisible to human vision) are acquired. All the images identified as RGB images in this report are derived from the spectral data sets and not from conventional color sensors.

The most common computational tool applied to MSI data for separating marker fluorescence from autofluorescence in biological samples is linear unmixing (18). This technique allows the segmentation of a MSI data set into a series of monochrome images that ideally reflect the concentration and distribution of the components generating the individual spectra used in the unmixing procedure. Effective use of such spectral tools to remove autofluorescence signals requires accurate spectra of the individual components. Frequently, however, the measured emission spectra in vivo may be different from those measured in vitro due to combinations of effects of the local environment, absorbance, and scattering. Both manual and automated tools have been developed that can determine the spectral shape of the fluorophore(s) in situ, as mixtures, without physically isolating them (16). This can be done in the same animal model being used in the study, thus requiring only a single positive and a single negative control sample and a few minutes of data acquisition and processing.

MSI methodologies can be used to image practically any of the fluorophores used in biomedical research, as long as the emission wavelengths of the dye are between 500 and 950 nm. These fluorophores can be ones xenografted into either a tumor implanted in an animal or an animal itself, they can be covalently bound to antibodies, peptides, or other agents that bind in an animal, or they can simply be fluorescent compounds that are injected into an animal. An example of antibody-targeted spectral imaging and analysis can be found in the report by Gao et al. (14), examining the distribution of quantum-dot-labeled antitumor antibodies in mice.

Results

Mouse Injected with a Fluorescent Agent

Mice with a shaved and depilated region were imaged before and after injection with a naturally fluorescing photodynamic therapy (PDT) agent that accumulates in a tumor after a short time. The PDT agent, with a 680-nm peak emission spectrum, was developed by Allan Oseroff at the Roswell Park Cancer Center, Buffalo, NY. Each mouse had subdermal tumors growing on its flank, whose location was circled with a black felt pen prior to the experiment. Figure 3 shows monochrome (680±10 nm) images of the pre- and post-injection mice. A small amount of increased brightness can be seen inside the marking circle of the post-injection animal. The fluorescence spectrum of the animal prior to injection is shown in the dashed green line in the graph at right in Figure 3. The region from which the spectrum was taken is shown by the green arrow. The spectrum shown in red at right is taken from the region over the tumor in the post-injection animal, clearly showing a large fluorescence signal with an emission maximum at 680 nm.

Figure 4 shows the results of the unmixing of the PDT agent fluorescence signal from the autofluorescence signal. The upper left image is an RGB representation of the fluorescence of the post-injection mouse. A small amount of red signal can be seen inside the region demarcated by the black line. The upper right image shows the unmixed autofluorescence signal, and the lower left image shows the unmixed PDT agent image, with the fluorescence from the dye clearly visible against a black background. The lower right image shows an image formed by combining the autofluorescence image, which had been left as a grayscale image, and the PDT agent image, which had been colored red for visibility. The red PDT agent signal in the combined image is centered inside the black fiducial markings drawn around the tumor.

Mouse Implanted with Three Fluorescent Protein-Labeled Tumors

A nude mouse was implanted with three subdermal tumors, each labeled with a different fluorescent protein: GFP, red fluorescent protein (RFP), and mPlum fluorescent protein (19).

The upper left image in Figure 5 shows an RGB representation of the fluorescence of this mouse. The 680-nm food autofluorescence in the intestines can be clearly seen at the right of the animal. The GFP tumor signal shows as a fairly visible green region. The mPlum tumor signal is barely visible as a deep red band across the animal. The RFP tumor is not apparent in the RGB image even on close inspection. The spectra for the GFP (green), RFP (red), and mPlum-FP (magenta) are shown at right. Note the highly overlapped signals from the RFP and mPlum-FP. Not shown in the graph are the spectra for the autofluorescence from the skin of the mouse or from the food.

The lower left image in Figure 5 shows the results of unmixing the five signals (GFP, RFP, mPlum-FP, food autofluorescence and skin autofluorescence) from each other. The unmixed GFP image has been colored green, the RFP image colored red, the mPlum-FP image colored magenta, and the unmixed food image colored blue. The skin autofluorescence image is not included in this composite. All three of the tumors are now clearly visible in the image.

Discussion

Bioluminescence and fluorescence techniques are proving to be central to practical molecular in vivo imaging in small animals, for life sciences research and pharmaceutical drug development. Bioluminescence offers advantages of high sensitivity, but with relatively expensive instrumentations and certain timing and sample preparation requirements, such as requiring genetically engineered cell lines or animals to express luciferase genes. Fluorescence techniques have historically been substantially limited by autofluorescence, but offer advantages of flexibility of labeling strategies, and multiplexing. With multispectral image acquisition and linear unmixing techniques, the sensitivity possible with fluorescence techniques has become significantly closer to that achievable with bioluminescence, thus enabling scientists to benefit from its flexibility, multiplexing capabilities, and lower cost instrumentation and reagents. Signal-to-noise considerations can be used to guide the choice of sensor, which in many cases will require only modest cooling capabilities, since the major noise sources will be shot-noise (signal-based) rather than arising from the sensor itself. With appropriate software tools for generating accurate spectral characterizations of the target and autofluorescent species, at least five spectrally distinct fluorescent sources can be unmixed in a single animal subject.