Spectral Imaging and Linear Unmixing
Over the past decade, a wide spectrum of high-performance fluorophores have been developed for investigations in fluorescence microscopy using advanced techniques such as laser point-scanning confocal, spinning disk, multiphoton, and total internal reflection. Among the advanced probes that are now available are genetically-encoded fluorescent proteins, semiconductor quantum dots, hybrid systems consisting of membrane-permeant synthetic fluorophores that target protein fusions, and stand-alone synthetics that exhibit a broad range of physical and spectral properties. These reagents are capable of targeting virtually any protein or peptide in living or fixed cells and many are also quite useful as indicators of biological dynamics. When used as single labels, the imaging of most fluorophores is straightforward and can be readily accomplished with standard illuminators and fluorescence filter sets. However, when using combinations of two or more fluorescent probes, the investigator should be aware of the potential for signal crossover in detection channels that occurs due to spectral overlap between fluorophores.
Advanced research microscopes are capable of imaging multiple fluorophores in a single experiment using single, dual, triple, or quadruple-band dichromatic mirrors as well as associated excitation and emission filter sets that are matched to the emission spectral profiles of the probes being analyzed. Unfortunately, the emission spectral profile of many synthetic fluorophores and fluorescent proteins spans a wavelength band ranging from approximately 50 to 150 nanometers, resulting in the potential for signal bleed-through of emission from one probe into the detection channel intended for another. The situation is complicated even further when the choice of fluorescent probes is limited or more than two fluorescent proteins are simultaneously targeted for investigation in the same specimen. Because the level of information afforded by studying the interaction between multiple probes in living and fixed cells significantly exceeds that available from using the same labels individually, methods have been developed to unravel complex mixtures of emission signals so that spectra can be individually resolved. Leading the field in this regard is a technique known as spectral imaging coupled to mathematically linear unmixing of the measured spectral profiles, which represents a synergistic union of optical imaging and molecular spectroscopy currently available as auxiliary hardware and software components on a wide variety of commercially available widefield and laser scanning confocal microscopes.
Illustrated in Figure 1 is the cutaway diagram of a high-performance Nikon A1 confocal microscope spectral detector unit that enables high-speed acquisition of multiple spectral profiles in a single scan. The spectral detector is interfaced to the scanning unit through an optical fiber that features wavelength resolution independent of confocal pinhole diameter. Fluorescence emission light entering the detector is first passed through a proprietary Diffraction Efficiency Enhancement System (abbreviated DEES) that separates incoming non-polarized emission light into two orthogonal polarized light wavefronts (termed p and s) using a polarizing beamsplitter. The purpose of the DEES system is to increase the efficiency of light diffraction by the gratings used to separate fluorescence emission into component wavelengths. After leaving the beamsplitter, the p-wavefront is rotated by 90 degrees (into an s-polarized wave) using a prism system and both beams are then diffracted by one of three interchangeable gratings. The diffraction gratings, which can be precisely controlled to ensure a high level of reproducibility, have wavelength resolutions of 2.5, 6, and 10 nanometers. Using the 10-nanometer grating, spectra over a range of 320 nanometers can be obtained in a single scan. Spectrally distinct wavelengths diffracted by the grating are focused by two independently adjustable cylindrical mirrors onto a 32-channel multi-anode photomultiplier such that the focused spots arising from each beam (the normal s and rotated p waves) overlap perfectly. A specialized shielding mechanism enables simultaneous excitation of mixed fluorophores by up to four lasers.
Among the many applications that benefit from spectral imaging coupled with software-based linear unmixing algorithms are those that require the ability to decipher the individual spectral profiles for a large variety of absorbing dyes and fluorophores in situations where a significant degree of spectral overlap occurs. Investigations involving live-cell imaging (in both animals and plants), immunofluorescence, karyotyping, clinical pathology, flow cytometry, in vivo imaging, and drug discovery can be enhanced using spectral imaging methodology. The spectral information available using this technique can often distinguish between measuring artifacts, such as autofluorescence, refractive index fluctuations, unsuspected fluorophore interactions, and environmental inhomogeneities, and obtaining the desired signal from probes that are the target of the experimental protocol. Aside from the utility in removing unwanted autofluorescence that can obscure detail in many specimens, spectral imaging is also a significant advantage in separating the overlapping emission spectra of fluorescent proteins and other fluorophores in dynamic fluorescence resonance energy transfer (FRET) experiments, which are often complicated by the requirement for exceedingly fast image capture.
Depending upon the subtle details of molecular architecture and elemental composition, the fluorescence emission spectrum of any particular fluorophore can be distributed over a broad wavelength range that varies between 30 and 200 nanometers. To demonstrate this concept, the "generic" absorption and fluorescence emission spectrum generated by a hypothetical fluorophore is presented in Figure 2. The spectral profiles illustrated in this figure have characteristic features that are common to all fluorophores, with the emission profile approximating (but not exactly) a "mirror image" of the absorption profile. In the shorter wavelength regions of the spectral profile, the emission spectrum features a steep increase in quantum yield as the curve approaches the maximum value (termed the peak emission wavelength). Following the peak, as the wavelengths continue to increase, the slope of the spectral profile declines far more slowly until it reaches a minimum value upon returning to the baseline. The bandwidth of fluorescence emission is generally measured by the width of the spectral profile at 50 percent of the maximum quantum yield and is often referred to as the full-width at half maximum (FWHM; Figure 2). However, as can be observed from the profile in Figure 2, the amount of fluorescence emission that arises in the longer wavelengths outside this region (in some cases exceeding 100 nanometers) can be significant.
When imaging with a single fluorophore in fluorescence microscopy, a longpass emission filter can be used to gather the maximum amount of emission over a broad spectral range. In such a case, there is no concern about signal interference (often termed crossover, bleed-through, or crosstalk) arising from emission overlap by another fluorophore. However, when imaging multiple fluorescent labels simultaneously, the emission profiles often share the same spectral region, especially in the longer wavelengths, and require restricted-wavelength bandpass emission filters that are centered near the peak wavelengths of the individual fluorophores to limit, or preferably eliminate, the unwanted detection of emission from other fluorophores. Even when using narrow bandpass filters, emission from one fluorophore can bleed into the detection channel intended for another. This problem is less severe when using synthetic dyes or quantum dots that feature narrow emission spectral profiles (typically 30 to 60 nanometers), but becomes far more serious when imaging two or more fluorescent proteins, which often have emission spectra that span hundreds of nanometers. Note that the width of the entire visible spectral region is limited to approximately 300 nanometers (from approximately 400 to 700 nanometers). Therefore, simultaneous imaging of two well-separated fluorescent proteins, each having emission spectral profiles spanning 150 nanometers, will result in detectable signal over a significant portion of the visible light spectrum.
The level of fluorescence emission signal that can be detected is determined by the width of the emission filter passband and the efficiency of fluorophore excitation, a factor that is determined by the absorption spectral profile of the fluorophore coupled to the passband width of the excitation filter. As a consequence, even in cases where fluorophores have significant overlap in their emission profiles, they may not produce detectable crosstalk if the excitation filter wavelength passband characteristics are carefully chosen. In addition, due to the fact that excitation efficiency sharply decreases at wavelengths longer than those surrounding the absorption peak (see Figure 2), fluorophores having the longest wavelength absorption properties can often be specifically excited without simultaneously exciting fluorophores that absorb at shorter wavelengths. For example, when imaging enhanced cyan (ECFP) and yellow (EYFP) fluorescent proteins (having highly overlapping emission spectra; see Figure 3(b)) crosstalk can be minimized by first exciting and detecting EYFP, followed by imaging the shorter-wavelength ECFP using excitation and emission filters designed to minimize excitation of EYFP. This technique relies on sequential imaging using bandpass emission filters that often can detect only around 50 percent or less of the available photons emitted by each fluorophore. Furthermore, such an approach will not be efficient when using fluorophores (such as ECFP and EGFP; see Figure 3(c)) that have too much spectral overlap or unusual Stokes shifts.
Illustrated in Figure 3 are computer-simulated live cells expressing several combinations of two fluorescent proteins to demonstrate spectral crossover in fluorescence imaging. The cells are Indian Muntjac deer skin fibroblasts expressing ECFP in the nucleus (channel 1) and either EGFP, EYFP, or monomeric Kusabira Orange fluorescent protein (mKO) as fusions to beta-actin in cytoskeletal filaments (channel 2). The region of spectral overlap for each probe combination is indicated in the spectral plots as a gray region within the filter window. These fluorescent protein combinations demonstrate varying degrees of emission spectral overlap. The combination of ECFP and mKO (Figure 3(a)) demonstrates the least amount of overlap and the ECFP-labeled nucleus does not show substantial signal in channel 2 (the mKO channel). In contrast, the combination of ECFP and EYFP (Figure 3(b)) exhibits considerably more overlap and the nucleus becomes discernable in channel 2. Finally, the combination of ECFP and EGFP demonstrates the highest degree of spectral overlap (Figure 3(c)), with significant signal crossover between the channels.
Attempting to use three or more fluorescent proteins in a single experiment requires the use of even more highly restricted excitation and emission filter strategies. In most cases imaging of two or more fluorophores results in several undesirable consequences, including speed limitations due to the necessity for sequential imaging, reduced sensitivity as the result of smaller filter passband size, and more complex labeling strategies that are necessary to minimize spectral overlap. Gathering images sequentially requires more time than simultaneous imaging and results can be compromised by rapid specimen motions during acquisition. Furthermore, the fluorophore signal levels in living cells are often low, especially for specimens with sparse target abundance or those expressing at endogenous levels. As a result, detection using limited passband filters can be challenging. The signal-to-noise of unmixed data sets often exceeds that obtained from narrow bandpass emission filters because the entire 150-plus nanometer emission bandwidth of fluorescent probes can be samples as opposed to just a narrow 30-nanometer emission window. Finally, the fluorescent protein color palette is still rather limited and the broad emission profiles make it difficult to cleanly separate emission or else require specialized filter sets. Thus, in live-cell imaging where high speed acquisition is often a mission-critical factor in the success of an experiment, these consequences can have a severe impact on the results of an investigation.
Spectroscopy in Microscopy
Spectral imaging merges the disciplines of microscopy and spectroscopy into a combination that enables determining the both the intensity and spectral properties of a specimen for each pixel in an image. Recent technological advances in imaging have produced highly sophisticated digital cameras and point-source detectors (such as photomultipliers) capable of creating information-rich images with high spatial resolution and dynamic range from a variety of specimens under numerous contrast-enhancing techniques, including brightfield, phase contrast, and fluorescence microscopy. These detector advances also extend the ability to create suitable images of faintly fluorescent specimens whose low signal levels were previously lost in the noise floor. In contrast, spectroscopy is a well-established field that involves gathering and analyzing a quantitative collection of light intensity values over a defined wavelength band, which can encompass any portion of the electromagnetic radiation spectrum. In microscopy, spectroscopic imaging techniques are generally limited to wavelengths ranging from the near-ultraviolet to the near-infrared.
Atoms and molecules have an intrinsic energy-band structure that can be examined with spectroscopy. Briefly, the process of light absorption excites an electron from the ground state to a higher energy level from which it can relax through several pathways, including decay back to the ground state with the associated emission of lower-energy light (fluorescence). Energy levels are intrinsic properties of each molecule and thus provide a precise spectroscopic fingerprint for that molecule. In fluorescence, specific molecules (termed fluorochromes or fluorophores) are attached to the structure of interest and used as light sources for imaging. An important concept to note in fluorescence microscopy is that there is often a direct linear relationship between fluorophore concentration and the amount of fluorescence intensity, especially at low concentrations. This situation enables the quantitative analysis of fluorescence in cases where the necessary signal can be successfully segregated from saturation, photoconversion, and photobleaching effects that often tend to disrupt the expected linearity.
Unlike the case with fluorescence, in other forms of optical microscopy including brightfield, enhanced contrast transmitted modes (phase contrast; Hoffman modulation contrast, HMC; differential interference contrast, DIC), reflected light, and scattering, the specimen is illuminated with an external broadband light source and the detector measures that same light after it interacts with the specimen. In order to analyze spectral data, the spectrum emitted by the light source must be taken into account and the measured signal is usually not directly proportional to the concentration of the chromophores or absorbing species in the specimen unless it is first converted to optical density units in accordance with the Beer-Lambert law. Nevertheless, most cells and tissues stained with absorbing dyes and imaged using brightfield techniques are prime candidates for spectral imaging analysis.
In order to measure the spectrum of an absorbing dye, fluorophore, or complete specimen with multiple labels, the transmitted or emitted light is first dispersed into its component wavelengths and the intensity at each wavelength or a very narrow band of wavelengths is measured. The spectral resolution is dependent upon the bandwidth of each measurement and increases as the bandwidth of the sampled channels decreses. A variety of different techniques can be used to disperse light, and most of them have been applied (at least in prototype instruments) to microscopy scenarios. Among the most important characteristics to consider when measuring spectra are resolution, wavelength range, and dynamic range. Spectral resolution is determined by the closest wavelengths that can be distinguished from one another and is a critical parameter for highly accurate spectral imaging measurements. Spectral range refers to the total number of wavelengths (in effect, the bandwidth) in a particular measurement. Finally, the detection limit and dynamic range define the minimum level of signal necessary for conducting measurements and number of distinguishable levels in a particular measurement, respectively. All of these values can vary for each fluorophore or absorbing species as a function of the spectral profile.
Illustrated in Figure 4 is a typical set of spectral images acquired in sequential bandwidths of six nanometers spanning the wavelength range of 500 to 692 nanometers to generate a lambda stack (discussed in detail below) containing 32 images. The specimen is a culture of adherent human cervical carcinoma (HeLa line) cells in which DNA and RNA was stained using Acridine Orange and imaged with a Nikon A1 spectral confocal microscope system. Images (512 x 512 pixels) were recorded using the 32-channel multianode photomultiplier at 24 frames per second using 488-nanometer laser excitation. Such a high acquisition speed, which is of significant benefit in live-cell imaging, is made possible by advanced signal processing technology coupled with fast analog-to-digital conversion circuitry that operates in tandem with the photomultiplier.
Spectral Imaging Considerations
Sophisticated techniques for spectral imaging first emerged in the field of remote sensing due to the requirement for analyzing satellite images containing data arising from reflected, refracted, and scattered sunlight, as well as shadows, all in the same scene. Methodology for spectral analysis is designed to enable characterization of multiple frequencies from image datasets in order to differentiate objects observed in a variety of landscapes and terrains. Similar techniques are used for other applications ranging from examining the chemical composition of materials to elucidating mechanisms in the formation of stars and galaxies. Thus, by examining intensity fluctuations as a function of spectral frequency (wavelength) to correlate individual pixels with matching spectral information, new details can be uncovered that are obscured when simply analyzing single images captured using light of mixed frequencies.
In general, satellite and celestial datasets are more complex than those obtained from images of biological specimens labeled with one or two probes because the number of spectral classes is usually far greater. Regardless, the spectral signature of satellite images is similar in complexity to that obtained when examining a mixture of overlapping fluorescent proteins in living cells or a fixed tissue specimen labeled with several absorbing dyes (such as eosin and hematoxylin). During the past several years, the unique approaches applied to satellite imaging have been successfully migrated to analysis of biological specimens in widefield and laser scanning confocal microscopy for a number of applications including the elimination of autofluorescence artifacts, for detection of weak and convolved Förster resonance energy transfer (FRET) signals, in karyotyping of human chromosomes, and for unraveling co-localized fluorophores on a pixel-by-pixel basis. The resulting spectral information can be used to pinpoint the location of specific fluorophores and dyes with high spatial precision, and is also potentially capable of producing information about interactions between two or more probes.
Under routine circumstances, and where experimental protocols permit, traditional confocal and widefield imaging techniques can be successfully applied through the careful selection of fluorescent probes and associated filter sets, as well as by implementing multitracking scan strategies with the proper controls to produce reasonable separation of fluorophore signals. Unfortunately, the increased use of multiple fluorescent protein colors (with their associated high degree of spectral overlap) to monitor intracellular interactions often limits the choice of experimental parameters. Furthermore, in live-cell imaging using a single fluorophore, natural autofluorescence can significantly interfere with detection in channels where the most popular green-emitting fluorescent probes (such as enhanced green fluorescent protein) are visualized. This background noise problem can also be seriously compounded by extraneous fluorescence that is introduced through the use of fixatives or DNA transfection reagents. In situations where fluorescent probe spectra significantly overlap or autofluorescence is excessively high, spectral imaging coupled with post-acquisition image analysis using linear unmixing algorithms can be utilized to untangle mixed fluorescence signals and clearly resolve the spatial contribution of each fluorophore.
Spectral Image Lambda Stacks
Similar in concept to the optical section (or z-stack) obtained from thicker specimens using high numerical aperture objectives in laser scanning confocal or deconvolution microscopy, the lambda stack is a three-dimensional dataset that consists of an image collection using the same specimen field acquired at different wavelength bands, each spanning a limited spectral region ranging from 2 to 20 nanometers. In contrast, typical imaging scenarios in all forms of optical microscopy involve acquiring a single image (or a successive group of images in time-lapse experiments) over the entire wavelength response band of the detector. Thus, whereas traditional imaging yields an intensity value for every pixel in the image (I(x,y)), a spectrophotometer measurement provides only a single spectrum (I(λ)). The lambda stack merges these by providing a spectral value at each pixel (I(x,y,λ)) to create what can be considered either as collection of images in which each image is acquired at a different wavelength (or narrow band of wavelengths), or as collection of different wavelength values for each pixel location.
In order to better understand the lambda stack concept (also commonly referred to in the literature as an image cube or spectral cube), a single pixel location in the lateral image dimension (having coordinates xi,yi) can be examined along the wavelength (zλ) axis. As illustrated in Figure 5(a), the intensity and/or color of the pixel i changes as a function of fluorescence emission signal strength and wavelength, respectively, when monitored from one end of the lambda stack to the other. By plotting pixel intensity versus wavelength on a linear graph (see Figure 5(b)), the emission spectral profile of the particular fluorophore spatially located at pixel i can readily be determined. It should be noted that the accuracy and resolution of an emission spectrum obtained using this technique is a function of the number of lambda stack images gathered at distinct wavelength bands, the spectral width in nanometers of each wavelength band (shorter bandwidths produce higher resolution), the physical quality of the specimen under investigation, and the photon sensitivity (quantum efficiency) of the detector.
A real-world example of a lambda stack acquired on a laser scanning confocal microscope in living cells using three fluorescent proteins having overlapping spectra is presented in Figure 6. The fluorescent protein markers used in this experiment are enhanced green fluorescent protein (EGFP from jellyfish; emission maximum at 507 nanometers), enhanced yellow fluorescent protein (EYFP from jellyfish; emission maximum at 527 nanometers), and the monomeric version of Kusabira Orange (mKO, emission maximum at 561 nanometers), a high-performance probe developed from a naturally-occurring coral protein. In this case, the individual lambda stack images were scanned in 10-nanometer wavebands ranging from 480 to 640 nanometers (Figure 6(a)) to generate a total of 16 spectral sections for the fluorescent protein mixture.
The first image of the lambda stack reveals the spectral signature of the specimen in the emission range of 480 to 490 nanometers, while the second image contains emission data from 490 to 500 nanometers (see Figure 6(b)). Note that virtually all of the fluorescence emission in the first two lambda sections arises from the short-wavelength tail of EGFP alone with only a very minor contribution from EYFP in the longer wavelength section (490 to 500 nanometers). In the next two lambda sections (500 to 510 nanometers and 510 to 520 nanometers), the contribution from EYFP steadily increases as the emission from EGFP reaches a plateau. In the three lambda sections between 520 and 550 nanometers, The EGFP signal begins to decrease as the contribution from EYFP emission reaches a maximum at approximately 530 nanometers. Likewise, the emission contribution from mKO becomes more significant in the band between 540 and 550 nanometers. Thus, in the 550 to 560 nanometer band, the relative contributions from the fluorescent proteins are approximately 10, 25, and 65 percent respectively for EGFP, EYFP, and mKO.
Emission contributions from EGFP and EYFP become diminished in the final wavelength bands (560 to 640 nanometers) as the emission from mKO dominates. In review, the wavelength bands at the extremes of the lambda stack (between 480 to 500 nanometers and between 590 to 640 nanometers) feature emission contributions that are dominated by the shortest and longest wavelength-emitting proteins, EGFP and mKO, respectively. Those wavelength bands in the center of the lambda stack (500 to 590 nanometers) contain fluorescence emission that represents some contribution from all three fluorescent proteins. As will be discussed below, the distribution of the mixed emission signal across the wavelength bands of the lambda stack can be linearly unmixed using reference emission spectral profiles from each probe to clearly separate the contribution of the individual fluorescent proteins.
Spectral Imaging Techniques
The principal instrumental consideration in optical microscopy for spectral imaging includes the ability to accurately segregate fluorescence emission or source light not absorbed by the specimen into its component wavelengths using a dispersive element or similar technique. A number of different methodologies have been implemented to generate lambda stacks in widefield and confocal microscopy using several detector designs. The most useful instruments have proven to be laser scanning confocal microscopes equipped with spectral detectors that rely on prisms or diffraction gratings to disperse fluorescence emission, which is then directed to one or more photomultipliers. Advanced confocal instruments contain a prism or diffraction grating to disperse the emission beam into its component spectrum, which is then passed to either a multi-anode photomultiplier that can simultaneously detect up to 32 individual channels of spectral information, or to slits that pass selected wavelengths to one detector and reflect shorter and longer wavelengths to additional slits. Widefield instruments utilize interference filters, acousto-optical tunable filters (AOTFs), liquid crystal tunable filters (LCTFs), interferometry, prisms, prisms coupled to reflectors, or gratings to generate lambda stacks for image analysis.
Among the many techniques that have been used to generate lambda stacks, the so-called wavelength-scan methods represent perhaps the simplest approach. In practice, a series of narrow bandpass interference filters (usually 5 to 20 nanometers wide) is used to gather a stack of images of the viewfield with each filter. Alternatively, a combination of shortpass and longpass filters having particularly sharp cut-off wavelengths can be used instead of bandpass filters. This type of configuration was used in several early confocal microscope implementations for spectral imaging, but has since been supplanted by more advanced techniques. When using filters, the bandpass size determines the number of wavelengths included in each lambda scan, and therefore, the spectral resolution. The emission filters are placed in a filter wheel positioned between the specimen and the detector to acquire successive images of the same specimen field after the wheel rotates new filters into place (see Figure 7(a)). In general, filter-based spectral imaging techniques are practical only in cases where a limited number of wavebands are required because the specimen must be repeatedly scanned with each filter, which can often result in excessive photobleaching. Furthermore, gathering lambda stacks with filters is a relatively slow process (requiring at least several minutes) that is unsuitable on the timescales necessary for live-cell imaging.
A more convenient method for obtaining wavelength-scan lambda stacks is to use variable filters (see Figure 7(b)), which can be finely tuned, offer a greater number of wavelength bands, and are generally more compact than filter wheels. The most widely used variable filter configurations are based on variable-spectrum interference filters and electrically-tunable optical filters. Circular variable filters contain interference filters that spatially alter the wavelength passband depending upon where the incident light passes through the filter. In practice, the filter is placed in the same location as a filter wheel, and is rotated to alter the passband. No commercial instruments are offered using variable filters, probably due to the fact that lambda stack images contain a minute superimposed spectral gradient that is manifested by filter design. However, several spectral detectors are available based on AOTF and LCTF designs (see Figures 7(c) and 7(d)) that position the tunable filter between the illumination source and specimen (when using broadband sources such as arc-discharge lamps) or between the specimen and the detector. Unfortunately, placing an AOTF or LCTF between the specimen and detector can result in poor transmission efficiency due to polarization and scattering artifacts.
The primary benefit of AOTF and LCTF filters is that they are electro-optical components with no moving parts and are capable of fast switching times when compared to filter wheels and slit-based systems. Liquid crystal tunable filters operate by transmitting a narrow band of wavelengths when a voltage is applied to a polarizable liquid crystal mounted between two linear polarizers. In most implementations, several stages are necessary to achieve high resolution spectral separation, a condition that also reduces the total amount of light transmission within the filter passband. When transmitting non-polarized light (as in fluorescence emission), a standard LCTF exhibits a light throughput of approximately 40 percent. If the non-polarized light is first passed through a polarizing beamsplitter, and an orthogonally polarized LCTF is used for the reflected and transmitted light, the efficiency can be doubled. The filtered light can then be detected with two separate detectors or recombined with another polarizing beamsplitter into a single detector, with the images placed either side-by-side or overlaid (the latter requires precise registration). Utilizing both polarization images enables the calculation of a fluorescence polarization anisotropy image, which can be particularly useful when conducting FRET measurements with fluorescent proteins.
Acousto-optic tunable filters employ a specialized crystalline compound, such as tellurium dioxide, which responds to acoustic waves by deformation of the crystalline lattice. At each acoustic frequency, the crystal deforms to produce a diffraction grating having a specific period that transmits a different wavelength (or narrow band of wavelengths). Both filter designs generate a lambda stack by capturing successive images at different wavelength bands and have the advantage of very high spectral resolution. In addition, the operator can choose an optimal exposure time for each separate wavelength band. On the downside, AOTF filters can be problematic, as discussed above, when used in the emission optical train (between the specimen and detector) due to poor light throughput, as well as image blur and shift artifacts. Several manufacturers have produced image quality AOTF devices. An AOTF can also be used as an acousto-optical beamsplitter (AOBS) in confocal microscopy to replace the usual laser combining filters. The AOBS enables much faster wavelength switching than filters and can be especially useful when combined with a super-continuum white light laser.
Spectral imaging time-scan methodology is based on acquiring a dataset that represents a superposition of the spectral and spatial information, but also requires a mathematical transformation of the gathered data to derive the spectral image. In practice, the technique does not require filters or prisms and is usually implemented in widefield microscopes by coupling an interferometer unit (Figure 7(e)) to a fluorescence microscope to conduct Fourier transform imaging spectroscopy. Interferometric methods split the incoming beam of light into two separate paths and introduce an optical path difference with a time delay between the two resulting beams. Upon arriving at the detector (CCD or photomultiplier), the two light beams are able to interfere. By measuring the intensity as a function of the optical path difference, an interferogram is created that is specific to the spectral properties of the specimen. The original spectrum can then be derived by applying a Fourier transformation algorithm to the interferogram. The most significant advantage of imaging spectroscopy is that the intensity at each wavelength is collected throughout the experiment and resolution can be modulated simply by adjusting acquisition parameters. The only disadvantage is that the entire spectrum of a specimen must be collected even in cases where only a few data points are required. Imaging spectroscopy is generally conducted on custom-built instruments due to the lack of commercial instrumentation.
The rapidly advancing technique of flow cytometry can also benefit from spectral imaging techniques. Due to the spatial limitations of flow cytometry when imaging, spectral imaging is conducted by selecting a smaller region of interest (usually having the dimensions of a single cell) and restricting the number of wavelength bands that are gathered. Thus, the fluorescence emission is dispersed by a diffraction grating or custom optical element and projected onto the surface of a time-delayed integration (TDI) CCD with the sensor pixel clock synchronized to the flow rate. A similar methodology utilizes computer tomography where a holographic dispersion element projects spectral and spatial information onto an area array CCD. In general, CCD camera speed is a major limitation for spectral imaging with the collection of a single lambda stack requiring several minutes or more. However, the emergence of faster systems can help overcome these pitfalls. Digital cameras also require repetitive scans to capture a lambda stack, which can lead to increased photobleaching and phototoxicity. The speed situation is far more critical in live-cell imaging where labeled structures can change spatial location during the acquisition of a lambda stack that consumes several minutes.
Spectral Imaging in Laser Scanning and Multiphoton Confocal Microscopy
By applying spatial-scan spectral imaging techniques, the entire spectrum of a specimen can be simultaneously acquired in a single pass using point-scanning or line-scanning instrumentation. This methodology is especially useful for live-cell imaging and for probing thick tissues where the specimen often must be repeatedly scanned and therefore exposed to large amounts of potentially damaging excitation illumination. Spatial-scan approaches to spectral imaging require dispersion of fluorescence emission using a diffraction grating (Figure 8(a) and 8(c)) or prism (Figure 8(b)) and have been widely implemented in commercial laser scanning and multiphoton microscopes. These instruments operate by separating fluorescence emission into its component wavelengths by passage through a prism or dispersion by a grating, followed by gathering selected portions of the spectrum using a variable-width slit or multi-channel photomultiplier. In order to control wavelength selection bandwidth in confocal instruments equipped with a slit, the aperture size is adjustable. For instruments containing multi-channel photomultipliers, the diffraction grating size can be altered (by rotating a new grating having different line spacing into the optical train) to control the number of wavelengths entering each channel in the detector. Common to both instrument designs is the presence of multiple detection channels that ensure the lack of physical gaps in the captured spectrum. Regardless of the bandwidth size, the number of images that can be gathered in a single lambda stack is limited only by the number of photomultipliers or channels.
The most versatile confocal microscope configuration for spectral imaging can dramatically enhance the acquisition speed of gathering lambda stacks by utilizing a multiple-channel photomultiplier to gather limited-size wavelength bands of fluorescence emission after it has been dispersed using a diffraction grating (see Figure 8(c)). This acquisition strategy has been successfully implemented in the Nikon C1si and A1 confocal instruments, each of which are capable of high-speed spectral acquisition with only a single scan. The multi-channel photomultiplier (often termed a multianode photomultiplier) in these instruments contains a linear array of individual 10-nanometer detection channels built into a single unit, which enables multiple emission bands to be imaged in parallel, thus severely limiting specimen photobleaching and phototoxicity. The Nikon spectral detection units feature several diffraction gratings with sampling increments of 2.5, 5 (or 6), and 10 nanometers that can be individually rotated into the optical path to adjust the spectral bandwidth of lambda sections. The dispersed emission is then passed to precisely defined channels in a 32-channel multianode photomultiplier to generate a separate image from each channel. The total bandwidth of fluorescence emission is dictated by the diffraction grating sampling increment: 2.5-nanometer sampling produces an 80-nanometer bandwidth, a 5-nanometer grating produces a 160-nanometer bandwidth, while the 6-nanometer grating generates a 192-nanometer bandwidth, and the 10-nanometer grating yields a 320-nanometer band. In the Nikon instruments, the spectral imaging detector uses a laser shielding mechanism that eliminates reflected laser light from the excitation source, and the diffraction grating can be tilted to select any of the sub-sampled bandwidths.
Among the advanced features of high-performance spectral imaging confocal microscopes are Nikon's unique proprietary diffraction efficiency enhancement system (DEES), which is designed to eliminate polarization artifacts, decrease wavelength losses at the grating, and capture the maximum amount of fluorescence emission. The DEES system operates by passing non-polarized fluorescence emission through a polarized dual-beamsplitter optical element to generate two component wavefronts termed p and s that are oriented parallel and perpendicular to the plane of incidence, respectively. The most proficient diffraction efficiency is observed with s-polarized light, so a polarization rotator is positioned in the pathway of the p-polarized light to generate s-polarized light, dramatically improving the efficiency of the grating system. As illustrated in Figure 9(a), the diffraction efficiency of p-polarized light is above 90 percent over the wavelength range of 450 to 675 nanometers. In contrast, the efficiency of s-polarized light is 80 percent at 450 nanometers and drops almost linearly to approximately 45 percent at 675 nanometers. Therefore, the Nikon DEES system can significantly improve light throughput, and therefore sensitivity, in the spectral detection unit. In cases where the spectral width must be adjusted, additional specimen scans can be conducted or adjacent detector channels can be combined (termed binning) to double, triple, or quadruple the width of the detection band.
Although slit-based spectral imaging confocal instruments are capable of imaging emission spectra at high resolution, they are relatively slow when compared to microscopes equipped with multianode photomultipliers. Even those instruments that feature mirrored slits to reflect a portion of the bandwidth to a second or third photomultiplier still suffer from a lack of imaging speed on the timescales necessary for live-cell imaging. In many cases, measuring a spectrum over 200+ nanometers in a slit-based system can take several minutes or more, thus hampering spectral imaging of specimens that undergo temporal motion throughout the imaging period. Among the advanced features that improve the performance of spectral imaging microscopes are sensitivity correction in multianode-based microscopes (see Figure 9(b)). These instruments are corrected for wavelength accuracy for each individual channel using emission lines and luminosity adjustment based on a traceable light source. Additionally, the ends of fiber optic elements and the detector surfaces are coated with proprietary anti-reflection agents to reduce signal loss and to achieve high optical transmission. Finally, advanced dual integration signal processing (DISP) technology has been added to the image processing circuitry to improve electrical efficiency, preventing signal loss while the digitizer processes pixel data and resets. As a result, the signal is monitored for the entire pixel dwell time, resulting in a dramatically improved signal-to-noise ratio. In fact, these combined technologies enable 32-channel spectral imaging (512 x 512 pixels) at speeds of 24 frames per second, fast enough for a wide variety of live-cell imaging applications.
In addition to performing spectral imaging using fluorescence emission to generate lambda stacks, the technique can also be expanded to utilize the excitation spectral properties of the fluorophores under investigation. Excitation-based lambda stacks can be acquired by varying the excitation wavelength accompanied by collecting fluorescence emission using a single detector. Due to the fact that emission is gathered in a single channel, the signal-to-noise ratio is typically high and presents an advantage for data processing. Excitation lambda stacks are analyzed using identical linear unmixing algorithms designed for emission spectral data (see below). In general, a broadband laser source capable of generating a wide spectrum of lines with fast switching is required for collecting high resolution excitation lambda stacks, but simpler forms can be implemented using several individual lines from different lasers. The primary disadvantage of excitation-based spectral imaging is the relatively slow imaging speed that arises from the requirement for sequential wavelength scans. The technique is therefore of more utility when imaging fixed cells and tissues rather than for live-cell imaging. In the future, as supercontinuum white light lasers are added to confocal microscopes, spectral imaging using excitation stacks should become more mainstream.
Spectral imaging is emerging as a powerful analysis tool in multiphoton microscopy, where the excitation source is usually a continuously tunable near-infrared pulsed laser. The potential ability to effectively separate fluorophores with multiphoton techniques is aided by the fact that many fluorophores with highly overlapping emission spectral profiles have distinct multiphoton excitation spectra with significantly less overlap. In such cases, linear unmixing should have the ability to separate fluorescent probes that would otherwise have too much emission overlap to be resolved. This prospect is particularly important when examining fluorescent probes having very similar emission profiles, such as Alexa Fluor 488, fluorescein, and SYTOX Green, which have far less spectral overlap in their excitation profiles than in their emission profiles. Multiphoton techniques are also potentially useful in separating emission of selected fluorophores from autofluorescence in the blue and green spectral regions.
Recent advances in the implementation of variable bandpass detection and spectral discrimination in laser scanning confocal microscopy offer far greater flexibility than traditional interference filters in fine-tuning the emission detection bandwidths for general imaging. In many cases, the successful separation of fluorescence emission is hampered by the fact that the instrument is not equipped with the optimum filters for the chosen fluorophores. Each of the spectral imaging configurations discussed above provides the ability to freely configure the detection wavelength range, which enables the investigator to design custom bandpass settings for virtually any fluorophore of interest. For example, using a Nikon C1si or A1 confocal instrument, the V-Filtering software option is capable of binning up to four selected channels to generate an image from one or more desired emission wavelength ranges (see Figure 10). This can be done regardless of whether the data is ultimately destined for linear unmixing analysis. Thus, modern spectral imaging confocal microscopes offer a significant advantage for eliminating fluorophore spectral overlap during routine imaging scenarios, and also provide the capability to easily create custom emission bandpass configurations for new fluorescent probes are they are developed.
Processing Spectral Images
The typical lambda stack acquired from a widefield or laser scanning confocal microscope generally contains hundreds of thousands or even millions of individual spectra (depending upon the image dimensions), one for each pixel in the collection. The resulting datasets are extremely large and therefore too complex to interpret visually. A comprehensive set of software tools is required for processing and displaying spectral image data, and this need has been addressed by a number of algorithms that have been published in the literature. Furthermore, all of the commercially available spectral imaging confocal microscopes are accompanied by advanced proprietary software packages designed specifically for data analysis and presentation using data acquired by the instrument. The analysis of spectral images can be performed based on either spectral features or image features present in the dataset (sometimes both), but most mathematical approaches involve algorithms that are collectively termed linear unmixing or linear decomposition. Software analysis of lambda stacks can be applied to virtually any combination of fluorescent probes, but image stacks composed of absorbing dye or reflected light spectral signatures must be converted to optical density before applying linear unmixing algorithms.
The original algorithms designed for spectral image analysis were developed primarily for the purpose of assigning individual signatures to objects captured in satellite images. The most useful mathematical approaches to this category of image analysis have been termed Principle Component Analysis (PCA), Supervised Classification Analysis (SCA), Multivariate Curve Resolution (MCR), and Linear Unmixing (LU). These algorithms are based on the assumption that the measured signal from each wavelength (or color) is linearly proportional to the percentage or concentration of that wavelength in the specimen. Such an assumption is generally correct when the concentration of absorbing dyes or fluorescent probes is low, but the results can deviate significantly from linearity as concentrations reach saturation levels. In that case, correction terms must be applied. Another important point to be clearly noted is that each fluorophore or absorbing dye has a unique spectral signature that can be independently determined for use as a reference in assigning the proper contribution from that probe or dye to individual pixels in a lambda stack. Gathering accurate reference spectra is a mission-critical step that should be carefully undertaken before the analysis of spectral images commences.
In a typical spectral imaging experiment based on fluorescence, there are usually several fluorophores present in the specimen, each one labeling a different structure. Throughout an image of this specimen, the fluorophores are found either separately or as a mixture depending upon their spatial distribution within the targeted organelles or macromolecules. The purpose of linear unmixing analysis is to determine the relative contribution from each fluorophore for every pixel of the image. In most cases, proper use of the algorithm requires recording of individual emission spectra in separately prepared control samples for all of the fluorophores used in the experiment. Control samples should prepared using the same techniques employed for the test specimen (such as mounting medium and cell type), and must be recorded using identical instrumental settings (gain, filters, objective, laser power, etc.) as the specimen under analysis. The importance of maintaining strict control over the preparation of samples and recording reference spectra cannot be overstated.
Another important criterion for linear unmixing of lambda stacks is that the separated spectra of all fluorophores must be distinguishable from one another and they must also be linearly independent such that none of the spectra can be produced from a linear combination of the others. This assumption is not trivial due to the fact that linearity criteria can potentially be disrupted by unknown or unintended interactions, such as energy transfer (FRET) between co-localized fluorophores, quenching, and environmental fluctuations. Considered an artifact in this situation, FRET can lead to a reduction of fluorescence emission intensity for the donor fluorophore accompanied by a slight change in its spectrum, along with increased emission intensity and potential spectral alteration for the acceptor fluorophore. In any event, however, the FRET effect is often very small, but should be considered when imaging fluorescent probes that have the potential to undergo FRET interactions. As a general rule of thumb, linear unmixing software performs best when using specimens that exhibit a high signal-to-noise ratio for all of the fluorophores that are being examined.
The fundamental concept underlying linear unmixing calculations is relatively simple. Each pixel in the spectral image is categorized as representing a mixture of fluorophore signals (intensities) when the measured spectrum (I(λ)) can be deconvolved into the proportion, weight, or concentration (C) of each individual fluorophore reference spectrum (R(λ)) when the values are summed. Thus, each reference spectrum of a pure fluorophore is described as Ri(λ) where i = 1,2,3.....n represents the index of the fluorophore (Ci). For a particular number of fluorophores (n), this relationship can be represented as:
I(λ) = C1•R1(λ) + C2•R2(λ) + C3•R3(λ) + ........ + Cn•Rn(λ)
Or more simply:
I(λ) = ∑i Ci•Ri(λ)
In practice, the signal intensity for each pixel (I) in the spectral image is determined and recorded during acquisition of the lambda stack and the reference spectra for the known fluorophores are measured independently in separate control specimens labeled with only a single fluorophore using identical sample preparation techniques and instrument settings. The overall spectral contributions from the various fluorophores in the specimen can then be determined as a simple linear algebra matrix exercise by calculating their individual contributions to each point in the measured spectrum, as described in the equations above. For many of the commercially available linear unmixing software packages, the solution is obtained by inputting reference spectral profiles and using an inverse least squares fitting approach that minimizes the square difference between the measured and the calculated spectra.
In order to ensure the best chances to obtain successful results when applying linear unmixing algorithms, several experimental criteria must be met. One of the most important considerations is to ensure that the number of spectral detection channels is at least equal to the number of fluorophores present in the specimen. Failure to meet this specification can result in multiple solutions to the spectral separation calculation and a unique result may not be possible. Another critical requirement for linear unmixing is that all fluorophores present in the specimen must be considered in the calculations or the results may be skewed towards the dominant (most concentrated) fluorophore at the expense of less concentrated species. Ironically, including spectra in the calculations that do not match any of the fluorophores in the lambda stack will not affect linear unmixing results (a zero contribution will be assigned to the missing fluorophore). Finally, autofluorescence and/or high background levels should also be defined spectrally (if possible) and treated as an additional fluorophore in order to achieve optimum results. Optionally, an error term can also be calculated and output as an error residuals image.
The linearity involved in adding fluorophore spectra is illustrated in Figure 11 for a mixture of two different, but highly overlapping hypothetical fluorophores having emission maxima residing in the yellow-orange (Fluorophore 1) and orange-red (Fluorophore 2) spectral regions. The black curves in Figure 11(a) through 11(c) represent the summed spectra of the two fluorophores at different concentrations: Figure 11(a) 1 to 1; Figure 11(b) 0.5 to 1; and Figure 11(c) 1 to 0.5. Although the spectra presented in Figure 11 represent examples of only three fluorophore combinations, the summed spectrum can readily be predicted for every possible combination of these two fluorophores simply by adding the intensities as a function of concentration. Note that the peak of the summed spectra changes with the proportions of the component fluorophores such that the maximum is 594 nanometers in Figure 11(a), 598 nanometers in Figure 11(b), and 589 nanometers in Figure 11(c). It should be emphasized that linear unmixing takes advantage of the entire spectral curve(s), not just the peak positions. Robust algorithms, such as those used in spectral karyotyping and confocal microscopy, also handle minute spectral shifts by sophisticated curve analysis and correction.
When analyzing the spectral content of a specimen labeled with two fluorophores, similar to that presented in Figure 11, the simplest approach is to match the summed spectrum from any particular pixel with all possible sum combinations residing in a spectral reference library. As an example, if the measured summed spectrum was a very close match to the black curve presented in Figure 11(a), it would indicate that the pixel contains a 50-percent contribution from each of the fluorophores and that they are evenly mixed in the specimen (at least for that pixel). Similarly, if the summed spectrum matches the black curve in Figure 11(b), one could assume that the pixel contains 66 percent of Fluorophore 2 and 33 percent of fluorophore 1. Thus, it can be summarized that linear unmixing operates by comparing a matrix representing the summed spectra measured in an image against a reference library of predicted spectra according to the best-fit parameters applied by the software. Once the spectral contribution from each fluorophore has been determined, the lambda stack can be segregated into individual images for each fluorophore, as illustrated in Figure 12.
Presented in Figures 12(a) and 12(b) are a pair of spectrally mixed and unmixed images, respectively, of an adherent culture of log phase Indian Muntjac deer skin fibroblast cells that were fixed in paraformaldehyde and labeled with SYTOX Green (nucleus), Alexa Fluor 488 conjugated to phalloidin (filamentous actin), and Alexa Fluor 514 conjugated to goat secondary antibodies targeting rabbit primary antibodies to PMP-70, a peroxisomal membrane protein (peroxisomes). A Nikon C1si confocal microscope was use to gather emission over the wavelength range of 470 to 550 nanometers using a 2.5 millimeter diffraction grating coupled to excitation using a 488-nanometer argon-ion laser (Figure 12(a)). Lambda stacks were linearly unmixed and pseudocolored (nucleus; red), (actin; blue), peroxisomes (green) to generate the final image shown in Figure 12(b). The brightfield images illustrated in Figures 12(c) and 12(d) were acquired using a fixed specimen of human liver tissue stained with eosin and hematoxylin. A Nikon 80i microscope equipped with a ChromoDynamics (Gooch and Housego) HSi Hyperspectral imaging detector and an Andor iXon EMCCD was used to capture a brightfield image of the mixed specimen (Figure 12(c)). After linear unmixing, specific regions of the specimen labeled with the two dyes are more clearly discernable (Figure 12(d)) by assigning pseudocolors.
Although spectral imaging and linear unmixing is becoming an important tool for analyzing complex mixtures of spectrally overlapping fluorescent probes in laser scanning confocal microscopy, this technique is also increasingly being applied to measurements conducted on pathological tissue and cell specimens stained with absorbing dyes and imaged using traditional brightfield microscopy. In order to conduct linear unmixing analysis on absorbing dyes, similar algorithms can be utilized after the data has been filtered to compensate for factors that apply to absorption rather than emission spectra. In contrast to fluorescence measurements, brightfield analytical techniques require that the absorption data gathered for each pixel must be separated from the spectral profile of the source light transmitted through the specimen. In effect, the transmitted light illumination must be measured as an additional reference. The absorption spectrum of a synthetic dye is linearly dependent on the concentration (as dictated by the Beer-Lambert Law). Mathematically, linear unmixing calculations for absorbing dyes are similar to those used with fluorescent probes, and can be expressed by the following equation:
A(λ) = ∑i εi(λ) • Ci • Li
where, as dictated by the Beer-Lambert equation, A is the absorbance of dye species (i), C is the concentration, and L is the specimen optical pathlength, which is usually measured in micrometers for sections of stained tissue. Prior to performing the calculation, the optical density must be determined for each absorbing species from the measured transmission values. It should be noted that conversion of transmission data to optical density can result in the introduction of significant noise levels when the signal level is low for a particular absorbing species, thus complicating unmixing results. In this case, spectral analysis is best performed by including only the absorbing region near the peak. Similar to the case for linear unmixing of fluorescent probes, concentrations for each stain in the specimen can be calculated for every pixel in the image if suitable reference absorption spectra have been independently determined.
In comparing brightfield absorption dye spectral imaging to that conducted in fluorescence mode, absorption dye specimens typically do not fade over time, whereas the very act of examining a fluorescent specimen can result in photobleaching. Specimens stained with absorption dyes typically have a dynamic range of 0.05 to less than 2.0 optical density (OD) units, with 0.05 OD being nearly indistinguishable from background, and 2.0 OD units corresponding to approximately 1 percent light transmission, which is relatively dark. Absorption images are also subject to distributional error due to glare if the condenser numerical aperture diaphragm is not adjusted correctly. Additionally, absorption dye imaging also requires a thorough understanding of the entire imaging system. For example, if the near-infrared light (heat) emitted by a standard tungsten-halogen lamp is not blocked (in effect, wavelengths greater than 720 nanometers), it can be detected by the CCD camera system and added to intensity counts as noise. Ideally, a wide bandpass filter with ultraviolet and near-infrared blocking should be included in the light path to remove unwanted wavelengths. Absorption dye imaging is further bounded by the fact that the maximum exposure time must be less than required to saturate the detector. In contrast, fluorescence exposure time is typically limited by detector thermal noise.
Spectral Imaging Applications
Spectral imaging provides the necessary foundation and tools to investigate phenomena in a wide variety of applications, including live-cell imaging, karyotyping, routine fluorescence imaging, drug discovery, detecting molecular interactions, and tissue pathology. The ability to gather partial or complete spectral information about the molecules being investigated enables the detection and differentiation of mixed fluorophores and absorbing dyes, even in cases where the probes exhibit similar color and highly overlapping spectral profiles. This powerful technique allows investigators to label multiple targets in cells and tissues with assurance that bleed-through and apparent co-localization will not interfere during image analysis. In addition, the information provided by spectral imaging coupled with linear unmixing can be used to distinguish between legitimate signals and artifacts produced by fixatives, transfection reagents, and mounting medium refractive index fluctuations. Spectral imaging is also becoming an important method for eliminating autofluorescence and for monitoring dynamic molecular interactions arising from resonance energy transfer.
Spectral karyotyping, based on the technique of fluorescence in situ hybridization, is one of the most popular applications for spectral imaging and has seen widespread acceptance. In a typical experiment, up to five different fluorophores are used to label each of the 24 human chromosomes, and the technique can also be applied to chromosomes from other species. Each chromosome is labeled with a different fluorophore combination, such as Cy5, FITC, rhodamine, one of the Alexa Fluor dyes, or any of the ATTO dyes. Combinatorial labeling results in 2N - 1 possible combinations. Thus, using a total inventory of only five fluorophores yields 31 possible two-dye combinations that can readily be discerned with analysis software (see Figure 13(a)). During analysis, the images are scanned for the spatial distribution of wavelengths, segmented, and then each pixel is classified based on a reference library of the five fluorophore spectra and a table of the known fluorophore combinations for each chromosome. The results of classification analysis are generally displayed in separate colors where each color represents a different chromosome (as illustrated in Figure 13(b)). The high specificity of the acquired spectral data enables a successful classification in most chromosome preparations, even when using complex tissue sections. By adding a sixth dye conjugates to probes for particular regions, or to probes for all the short chromosome arms, additional information can be obtained for automatic karyotyping. In practice, spectral karyotyping is often combined with DAPI banding.
The complex signals encountered in FRET microscopy are confounded by the excessive amounts of spectral overlap that are required by the fluorophores in order to undergo resonance energy transfer. Therefore, in addition to the ability of spectral imaging to separate fluorophore spectra in multicolor fixed and living cells, the technique is uniquely suitable for unraveling the emission contributions from donor and acceptor fluorophores in FRET imaging. Modern high-performance confocal microscopes equipped with multianode detectors are particularly suited for FRET analysis due to their high rate of image capture, which is often necessary when investigating fluorescent protein biosensors that operate on the millisecond timescale. With 32-channel multianode detectors, spectral imaging confocal microscopes can acquire the entire spectral response from both FRET fluorophores in a single scan. However, even though spectral imaging is capable of simultaneously detecting both fluorophore emission signals in FRET, the technique is incapable of distinguishing between acceptor emission generated through energy transfer and signal that originates from direct excitation. Therefore, proper controls using donor and acceptor proteins expressed separately are necessary for quantitative analysis. In cases where spectral imaging is used to evaluate FRET in fluorescent protein biosensors expressed as a single polypeptide, controls are less important.
Among the most critical artifacts that leads to a reduction in signal-to-noise during live-cell imaging is autofluorescence, which arises from a number of sources, including naturally fluorescent biomolecules (such as NADH, riboflavin, and elastin), DNA transfection reagents, culture media, and exogenous agents (drugs and biochemicals) added to the imaging medium. Fixative-induced autofluorescence is also particularly problematic when imaging cells and tissues that have been prepared with paraformaldehyde, and the artifact can interfere with whole body imaging in specific tissues (brain and skin). In addition, plant tissue tends to exhibit a high degree of intrinsic autofluorescence throughout the visible spectral region. One of the most powerful applications for spectral imaging is to eliminate autofluorescence from specimens labeled with weakly emitting fluorophores or those that have sparse targeting. In most cases, autofluorescence can be treated as a separate fluorophore having a distinct spectral profile (most readily determined in control samples) that can be unmixed from the signals of interest, and thus be reduced or eliminated completely from the final images (see Figure 13(c) and 13(d)). It is important to note that the level of autofluorescence is reduced at longer wavelengths in live-cell imaging, so careful choice of fluorophores emitting in the orange and red regions can help reduce this artifact.
Pathological specimens stained with multiple absorbing dyes and imaged under brightfield illumination are excellent candidates for spectral imaging and linear unmixing in widefield microscopy. Many of the common synthetic absorbing stains and dyes used to stain cell smears and tissue sections exhibit complex spectra that feature significant levels of overlap. However, in brightfield imaging the signal is usually strong and photobleaching is minimal or nonexistent. Spectral imaging can be used with brightfield specimens labeled with several dyes to yield separate images that present how the specimen would appear when stained with only a single dye. Post-processing can then be conducted to determine co-localization of dyes within specific structures. The most serious artifacts encountered with spectral imaging of brightfield specimens occur when imaging dyes that contain precipitates that scatter rather than absorb light or when cells and tissues are overstained.
Practical Aspects of Spectral Imaging
The technique of spectral imaging and linear unmixing has the potential to yield excellent results in situations where the experimental protocol is optimized to take advantage of the instrumentation capabilities and software parameters, ensuring that compromising artifacts are not unintentionally introduced. In short, the success or failure of most spectral imaging experiments is often determined by variables that are under the control of the investigator. The most important (and mission-critical) aspect is to obtain accurate reference spectra from the control fluorophore samples that faithfully represent the true spectral profiles. Additionally, great care must be taken to ensure that both control and test specimens are prepared under identical conditions with regards to fluorophore concentration, culture media, fixatives, wash buffers, mounting media, and optical quality of the glass slides and coverslips. Instrumental parameters should also be the same for controls and specimens under investigation. These include using the same objective, immersion oil, laser power, photomultiplier settings, pinhole diameter, dichromatic mirrors, wavelength scan range, and pixel dwell time. Under ideal conditions, and where the spatial distribution of fluorophores permits, reference images can be acquired in non-overlapping regions of the test specimen.
Even though advanced linear unmixing algorithms are capable of resolving the spectra of fluorophores that feature a significant degree of overlap, there are limits in the ability of most software packages to distinguish between fluorophores having spectra that are virtually superimposed. Such phenomena are a rare occurrence that is usually only seen in closely related fluorescent proteins and synthetic dye derivatives. For example, Alexa Fluor 488 and fluorescein are both xanthene derivatives having similar substituents with the major difference being that the former is sulfonated to increase solubility. The spectral peak of these two fluorophores is separated by only a single nanometer and the emission curves overlap almost completely. It is therefore impossible to distinguish between Alexa Fluor 488 and fluorescein using spectral imaging and linear unmixing. However, bear in mind that this is an unusual case and most common fluorophores used for labeling cells and tissues can be readily resolved using this powerful technique.
Matching as closely as possible the fluorophore concentrations and/or expression levels of fluorescent proteins in control and test specimens will help to ensure satisfactory results in spectral imaging. In general, investigators should strive to achieve the highest signal-to-noise levels as possible. Because individual channels in spectral imaging instruments range from approximately 2 to 10 nanometers in bandwidth, the sensitivity of the instrument is always limited by the number of photons that are able to register with the detector in each wavelength band. As a result, acceptable spectral separation of fluorophores at a resolution of 5 nanometers or less is only likely when using specimens that are labeled with bright fluorophores. Unfortunately, at very high resolutions (under 5 nanometers), typical biological specimens often exhibit poor signal-to-noise and image quality when labeled with fluorophores that are very dim or have sparse targeting. Thus, the best linear unmixing results are obtained when detection channel width is as large as possible.
Among the other experimental concerns in spectral imaging are high background levels, excessive detector and optical system noise, and autofluorescence. High background can occur from incorrectly targeted fluorophores, overstaining, laser line noise, mounting media inhomogeneities, immersion oil mismatches, stray light, and autofluorescence. In most cases, background levels can be reduced using subtraction techniques after the spectral data is collected. Detector and optical noise become problematic when imaging weakly fluorescent probes, but can often be effectively eliminated by reducing the number of detection channels and scanning over larger bandwidths. Autofluorescence is best dealt with by including it as a separate fluorophore channel having a distinct spectral profile during linear unmixing. In conclusion, careful attention to instrument configuration details, sample preparation techniques, and the choice of optimum fluorophores are the key to successful results in spectral imaging experiments.
George McNamara - Miller School of Medicine, 1450 Northwest 10th Avenue (R-134), University of Miami, Miami, Florida, 33136.
Jeffrey M. Larson and Stanley A. Schwartz - Nikon Instruments, Inc., 1300 Walt Whitman Road, Melville, New York, 11747.
Michael W. Davidson - National High Magnetic Field Laboratory, 1800 East Paul Dirac Dr., The Florida State University, Tallahassee, Florida, 32310.