Multi-photon excitation can be employed to excite the autofluorescence of naturally occurring fluorophores in tissues [2,4,6,7]. This affords non-invasive imaging of skin with subcellular resolution without adding dyes [2,4,6–8,12]. Autofluorescent components that can be observed include NADH [13], keratin [14], elastin [15], FAD [16], melanin [17,18] and in addition the SHG of certain types of collagen can be detected [10,19]. Images of the layered structure of skin can be easily recorded with a quality comparable to histological sections [8,17].

The combination of tissue imaging and spectroscopic detection such as fluorescence lifetime imaging [4], spectral imaging [20] or hyperspectral imaging [21], enhances image contrast and allows fluorescent compounds to be discriminated based on differences in fluorescence decay kinetics or emission spectra. Lifetime imaging methods typically utilize time correlated single photon counting [4] or time gating [22], spectral imaging can be performed using a set of emission filters [21,23] or dispersive optics and a PMT array [17] or CCD camera [20]. The advantage of the latter approach is that it affords detailed (high resolution) spectroscopic analysis; this was for instance demonstrated in in vivo mouse skin imaging experiments [7,8]. Rapid visualization of both spectral and morphological information can be achieved in real time by “real color” RGB representation of the spectral images [7,8]. Endogenous fluorophores can be identified based on their spectral signature [8] and spectral unmixing can be employed in an offline analyses to determine the relative contributions of individual fluorophores to the signal [7].

The disadvantage of conventional CCD camera based spectral imaging systems is their slow spectral acquisition rate. For instance, we previously reported a nonlinear spectral imaging microscope with a spectral readout of 500 Hz, resulting in a 2-minute image acquisition time for images of 224 × 224 pixels [20]. The background signal in these experiments was typically not constant over the image. To correct for this background, background spectra were extracted from low intensity regions-of-interest; an elaborate procedure that does not account for variations in the background within an image. The slow acquisition time makes the system sensitive for motion artifacts and the errors in the background correction complicate obtaining accurate spectral information. These limitations make the system less well suited for in vivo imaging.

Here, we optimized the image acquisition hardware and procedure of nonlinear spectral imaging for in vivo applications. The most important improvements are: (1) use of an electron multiplying CCD (EMCCD) based spectrograph that can acquire 8000 spectra/s, and (2) simultaneous acquisition and correction for the background signal. This approach proved to be fast and robust; it accurately corrects for stray light and temporal drift of the background. The emission is detected with high quantum efficiency by the EMCCD camera; at normal excitation power levels (typically 5-20mW) there is sufficient signal for RGB visualization and detailed spectral analyses of ROI’s.

The basic design of the spectrograph is simple: it consists of a fused silica equilateral dispersive prism (Linos, Göttingen, Germany), a focusing lens (f = 90 mm) and an EMCCD camera (Cascade 128 + EMCCD, Photometrics, Tucson, AZ, USA). The back illuminated CCD is equipped with on-chip multiplication gain, where the signal is amplified before the readout electronics. Therefore, this camera is very suitable for detecting low light levels. An EM gain of 30%-of-maximum has been used in all experiments. Spectral acquisition is controlled by software written in V++ (Digital Optics, Auckland, New Zealand). A low level camera readout script was written and a customized driver is used for optimized performance.

The high spectral acquisition speed is the result of the optimized read-out scheme of the EMCCD. The rate limiting step is the speed of A/D conversion. Therefore, the number of read out pixels is reduced and the collection of photons occurs simultaneously with the read out of the data of previous pixels. The spectrum is projected on to the EMCCD chip parallel to the read out register (see Fig. 1, right). Three sections of 10 rows of 128 pixels are defined. The emission spectrum is projected on the central section. The other two sections are used as background reference spectra. The readout script is synchronized with the pixel clock of the microscope. For each voxel of the acquired image, the sensor area of the camera is exposed until the next pixel clock trigger. After the trigger, the content of the sensor area is shifted to the frame transfer area of the CCD chip. While the emission and background spectra of the next voxel are collected in the sensor area, the three sections of interest are shifted from the frame transfer area to the serial read-out register. The 10 vertical rows of each section are binned and shifted to the first multiplication register. The spectra are transferred through four multiplication registers before being read out. Also the transfer of the spectra through the registers is optimized. During normal operation, the data of three voxels are simultaneously recorded on the EMCCD chip: one in the sensor area, one in the frame transfer area and one in the register. The resulting raw data file is 128 × 3 × N² pixels for an N × N voxel image.

Experimentally the minimum dwell time was found to be 85 μs per voxel. In all the experiments reported here, a pixel dwell time of 128 μs was used and all images are 224 × 224 pixels. This results in an image acquisition time of 6.5 seconds.

Wavelength calibration was carried out with great care. In the nonlinear spectral microscope, the emission from the focal volume of the objective is dispersed and imaged onto the CCD. Small shifts in the focal position affect the calibration of the spectrograph. This makes it very difficult to use an external light source for calibration purposes; multiphoton excited emission has to be used instead. Here, wavelength calibration is accomplished by inserting a multiband bandpass filter (Semrock FF01-390/482/563/640-25) in the emission path of the microscope. The filter has steep edges (~1 nm) and by dividing spectral recordings with and without the filter, positions of the edges are obtained. Calibration is finalized by fitting the pixel positions with a 2nd order polynomial. The correction for the wavelength dependent sensitivity of the system (flat-field) is composed of the spectral response of the camera multiplied by the transmission curves of the optical components in the emission path (e.g., dichroic and filter) and a correction for the wavelength dependent dispersion of the prism. The flat-field correction is constructed from data supplied by manufacturers and the material properties of the prism.

The above background correction procedure significantly improves the quality of the images and was applied to all the images shown here. A Z-stack of 10 XY images of in vivo Caucasian human skin is shown in Fig. 3 . In addition, an XZ image through the center of the stack is shown in the same figure (top left). The pixel dwell time was 128 µs per pixel, 6.5 seconds for a 224 × 224 pixel image and ~1 minute for the total stack. The excitation power was 15 mW, sufficiently low to avoid significant photodamage [25,26]. Various structures can be distinguished. A part of the Stratum Corneum, corresponding to a furrow, extends into the recorded stack and exhibits bright green autofluorescence that probably originates from clusters of keratin. Epidermal cells have predominantly blue/green emission because of NADH, keratin and flavins. Some cells are melanized, which results in additional yellow/green emission. In the dermis, violet second harmonic generation (SHG) of collagen fibers is observed, as well as blue/green autofluorescence of elastic fibers. Vessel-like structures with orange emission can also be distinguished in the papillary dermis (upper region of the dermis) and correspond to blood capillaries.

A major advantage of the high-resolution EMCCD based spectrograph is that full autofluorescence spectra of skin components can be recorded. This is demonstrated in Fig. 4 for Caucasian and Asian skin. The images were recorded at the junction between epidermis and dermis. The fluorescence spectra are averaged on four regions of interest (ROIs) that contain: (1) epidermal cells, (2) melanized epidermal cells, (3) collagen fibers and (4) elastin fibers. In the dermis, two major components can be distinguished: a second harmonic generation (SHG) band at 380 nm from collagen and autofluorescence that mainly originates from the closely located elastin fibers. Because of the overlap of the networks formed by these two major components of the dermis, there is some bleed-through in their spectra (Fig. 4). The spectra of the collagen fibers SHG and elastin fibers autofluorescence are the same for both volunteers. Our data indicate that the epidermal cells, however, contain a skin type dependent contribution of melanin. Already from the RGB images, it is clear that melanin contribution (yellow/green autofluorescence) is stronger in the Asian skin type. Since melanin is composed of heterogeneous oligomers of fluorophores, its emission spectrum strongly depends on experimental conditions (as shown for synthetic melanin by Teuchner et al [18,27]). Here, the spectra of melanin enriched cells exhibit an additional autofluorescence component at wavelengths longer than 550 nm. For the least melanized epidermal cells, the autofluorescence peaks at 480 nm, and the emission is a combination of NADH (450 nm), keratin (470 nm) and FAD (520 nm) autofluorescence. The spectra of nonmelanized epidermal cells of the two skin types are comparable. The melanized cells in Asian skin exhibit much stronger melanin autofluorescence than in Caucasian skin. For comparison, the spectrum of nonmelanized epidermal cells is added (gray line) to the spectrum in Fig. 4.

To confirm that the additional yellow emission is due to melanin, a series of images at the papillary epidermis/dermis interface were recorded (Fig. 5 ). Here, autofluorescence of melanin can be detected [28]. For three volunteers with Caucasian skin, the same observation is made: clusters of cells exhibiting yellow or green emission are located around dermal papilla. The difference in apparent color is likely due to different levels of tanning; low melanin levels appear green in the RGB images due to the presence of the red melanin fluorescence and blue autofluorescence (NADH, keratin). An increasing level of melanization results in a red shift of the observed color from green to yellow.

Improvements in acquisition speed reduce the amount of detected signal. Due to the use of an EMCCD in our setup this is partly compensated. EMCCDs have a high sensitivity and low readout noise even at fast readout speeds. The sensitivity is much higher than systems equipped with multiplier tubes (PMTs) and it extends well into the red. The effect of signal level is exemplified in Fig. 5, where the signal-to-noise (S/N) ratio is varied by averaging the spectrum in each pixel by binning (N = 1, 9 (3 × 3) and 25 (5 × 5)). Note that the spectra are averaged, and that the resulting R, G, and B values are scaled and multiplied by the original (not-averaged) intensity value in the pixel. Spectral averaging does not visually improve the color contrast in the images. The spectrum recorded in 1 pixel (128 μs, ~15mW, image acquired at the epidermal-dermal junction) is sufficient for RGB conversion. Since RGB conversion only involves multiplication of spectra, it is fast. On a standard computer, an image of 224 × 224 pixels is converted in less than a second.

The real color representation of the spectral images can be used to discriminate between different structures in tissue. To investigate the sensitivity of the real color RGB representation to noise and its uniqueness for the structural features in skin, ternary plots of the RGB values are made (see Fig. 6 ). These ternary plots display the relative contribution of the R, G, and B values for selected pixels. Five ROIs are selected: epidermal cells (EC), partly melanized EC, melanized EC, collagen and elastic fibers. The values of the pixels in these ROIs are overlaid in one plot. The ternary representation of unbinned pixels is not sufficient to uniquely identify components. By averaging over N = 9 and N = 25, however, clearly non-overlapping clouds of pixels are observed for most of the ROIs. The RGB values of the pixels containing elastic fibers are, however, very similar to the ones that represent partly melanized epidermal cells. This is not surprising since the spectra are very similar (see Fig. 4). Since they are located in a different layer of the skin, it is nevertheless easy to separate them.