Raman imaging is a powerful technique to reveal the chemical composition of samples. Due to its widespread availability and easy handling, confocal Raman microscopy is usually the method of choice for high demands on spatial resolution and optical sectioning, although near-field techniques such as surface enhanced Raman scattering (SERS) and tip-enhanced Raman scattering (TERS) [1,2] offer resolution in the nm regime. Recently, the concept of structured illumination has been experimentally transferred to far-field Raman microscopy, leading to a 1.4-fold improvement in the lateral directions [3]. The advantageous properties of confocal imaging are directly linked to the size of the detection pinhole or, in fiber-coupled systems, to the core diameter of the collection fiber.

Unfortunately, Raman Scattering yields a low signal compared to fluorescence imaging and, therefore, the collected light intensity is the limiting factor in most applications, especially in life science applications where the laser power has to be kept low to prevent sample damage. As a consequence, one often decides in favor of signal strength by choosing collection fibers whose core diameters are larger than the Airy disk, thereby sacrificing the confocal resolution advantage to a large extent.

In 1988, Colin Sheppard proposed using pixelated detectors in confocal microscopy, which, in conjunction with a post-processing technique denoted as “pixel reassignment” [4]—enable achievement of the highest resolution without discarding any light. The resolution of the final image is then determined by the size of an individual detector pixel (which here takes the role of the pinhole) and will asymptotically reach an improvement of $\sqrt{2}$ over the Abbe limit. An experimental realization of the concept was not shown until more than 20 years later [5], probably due to a lack of suitable detectors at the time, but has since triggered a large interest in the community and been applied to linear [6,7] and two-photon fluorescence microscopy [8]. In addition, all-optical reassignment setups without the need for any post-processing have been shown [9–11]. The technique has since become known as “image scanning microscopy” (ISM).

To the best of our knowledge, we present the first implementation of the ISM method to confocal Raman microscopy by minimally modifying the detection path of a commercial confocal Raman microscope (WITec alpha-300). The optics and detection system of the original microscope are left unchanged. Our specific implementation improves the lateral resolution by 28% compared to an image of equal brightness taken with the original microscope. The advantage can also be seen as a signal gain: compared to the standard confocal Raman microscope providing equal resolution, our implementation allows for roughly three times higher light collection efficiency.

In a standard fiber-coupled confocal Raman microscope [see Fig. 1(a)], the scattered light is collected and focused onto a fiber end, taking the role of a pinhole. The other fiber end is plugged into a spectrograph. The most common spectrograph type used in Raman imaging is the Czerny−Turner, where the light is first collimated by a curved mirror or lens, then diffracted by a grating and finally focused onto a detector. Usually, microscopes allow for an easy fiber exchange, so that the microscopist can choose between high resolution (small core diameter) or signal (large core diameter), or any intermediate case. A typical choice is a fiber whose core diameter roughly matches the size of the Airy disk formed by the light focused onto its end. This ensures good sectioning capability and high light efficiency. The lateral resolution, in this case, however, is degraded by about 30% compared to the theoretical confocal limit (obtained with a fiber core diameter close to zero).

 

Fig. 1. (a) Standard fiber-coupled confocal Raman microscope setup. (b) Our suggested modification, which consists of a hexagonal lenslet array and a matching round-to-linear fiber bundle. The linear end is plugged into the spectrometer and facilitates the collection of seven spectra in parallel. An additional magnification adjusts the system to its optimal performance. A closeup of the lenslet array is given in (c).

Download Full Size | PDF

Our modification of the standard setup is shown in Fig. 1(b). The single collection fiber is replaced by a bundle of seven optical fibers, which are hexagonally packed at the microscope end and linearly arranged at the detector end. Such “round-to-linear” fiber bundles are known to be useful for hyperspectral imaging [12], but usually they are used in spectroscopy to shape collected light into the form of a slit, thus allowing for improved spectral resolution. However, in contrast to the normal operation mode, where the sum over all fiber signals is recorded, we acquire an individual spectrum for each fiber. Here we exploit the fact that our charge-coupled device (CCD) detector (Andor Newton) provides many more lines than are actually used for operation with a single fiber. (Only approximately 10 of 200 CCD lines are normally used.) Thus, instead of using on-chip binning of all illuminated CCD lines, we only bin those lines that contain spectral information of the same fiber. For every spatial sampling point, we thus read out seven spectra from the CCD, each containing the signal delivered by an individual fiber [see Fig. 2(a)]. We would like to note that binning pixels is not absolutely necessary, as the summation could, in principle, also be done by post-processing, but this significantly reduces camera-related noise and readout time.

 

Fig. 2. (a) Camera frame at one scan point showing the seven acquired spectra is depicted. (b) Spectra at the same point in space with and without the pixel reassignment are shown. The lower panel (red curve) corresponds to the difference of the two spectra.

Download Full Size | PDF

To maximize coupling efficiency, we employ a hexagonal lenslet array in front of the round fiber bundle end. This prevents light losses caused by space between the individual fibers. Our lenslet array has a pitch of 200 μm (APH-Q-P200-R2.5 from Advanced Microoptic Systems), while the fiber bundle exhibits a fiber pitch of 220 μm (Thorlabs BFL200LS02). Although the two pitches should ideally be identical, the small mismatch in our system has no negative influence on the performance.

The effective entrance aperture of our detection system consists of seven hexagonally arranged lenslets and has a diameter of 600 μm, which is much larger than the size of any standard collection fiber of our Raman microscope (25–125 μm core diameter). Thus, an additional magnification of the focal spot is required. The simulations show that the highest resolution gain is achieved when the pitch of the lenslet array is close to the full-width at half-maximum (FWHM) of the local intensity focus (see Fig. 3). Therefore, the effective diameter for light collection is only about 30% larger than the Airy disk, and the sectioning capability is still maintained. For our experimental parameters (wavelength of 532 nm, $50\times 0.7\mathrm{NA}$ air objective), achieving this condition requires an additional 10-fold magnification. In contrast to the objective lens, which needs to meet high imaging quality standards, the demands on this second magnifying lens are much more relaxed, as it merely needs to magnify an on-axis light spot of about 50 μm at a very low NA of about 0.015. For this reason, it is possible to use a low-cost singlet lens (we use a standard aspheric fiber-coupling lens) without compromising the image quality. The light efficiency of our implementation is equivalent to that of a standard microscope equipped with a 60 μm collection fiber, while the spatial resolution of the final image is determined by the size of a single lenslet, corresponding to a 20 μm fiber.

 

Fig. 3. Simulation results showing the ideal pitch of the lenslet array in units of Airy disks to achieve the highest resolution gain, i.e., improvement in resolution compared to a standard, single-fiber-based microscope with the same light collection efficiency (red, solid curve) and the associated intensity gain compared to a single fiber yielding the same resolution (green, dashed curve); the inset shows a simulated image of the PSF at the position of the lenslet array.

Download Full Size | PDF

We investigate the imaging performance of our design by means of numerical simulations, which are based on separately deriving two-dimensional excitation and detection PSFs, i.e., $h_{\text{ex}}$ and $h_{\text{det}}$, followed by calculating the total imaging PSF $h_{\text{tot}}$ of every fiber as follows:

To achieve improved resolution, however, post-processing of the recorded data is required as described by Sheppard [4]. For this purpose, a Raman image is constructed from the data collected by each fiber by adding up the signal in the desired spectral band at every spatial scan point. The confocal Raman images obtained this way exhibit lateral shifts with respect to the image of the central fiber (see Fig. 4), defined by ${\mathbf{s}}_n={\mathbf{a}}_n/(2M_{\text{tot}})$, with image index $n$, which are determined by the lateral positions ${\mathbf{a}}_n$ of the corresponding lenslets at the entrance pupil and the total magnification of the system $M_{\text{tot}}$, i.e., the product of the magnifications of the used objective and the additional lens in front of the lenslet array.

 

Fig. 4. Confocal images constructed from the data collected by each individual fiber. Lateral shifts according to the fiber position have to be compensated for before summing the signals to achieve the resolution gain associated with ISM.

Download Full Size | PDF

The shift for the central fiber is zero as it defines the center of the coordinate system, while the positions of the remaining six fibers can be easily determined form the hexagonal array to be ${\mathbf{a}}_n=P_L·[\cos (n\frac{\pi }{3}),\sin (n\frac{\pi }{3})]$ for $n\in [1,6]$, where $P_L$ is the pitch of the lenslet array [see Fig. 1(c)]. Alternatively, the shifts ${\mathbf{s}}_n$ can be experimentally determined from cross-correlations between the images of the central fiber and the other fibers. In the final step of the post-processing, the sum over all the back-shifted images is calculated.

We conducted our experiments with seven fiber lenslet pairs but, in principle, the number of used detection channels can be larger. The number of 19 seems to be the next logical choice by accessing another “ring” of lenslets in the hexagonal array. This would enable a resolution gain of 1.35. The gain finally converges to a factor of approximately $\sqrt{2}$, corresponding to a perfect confocal system with an infinitely thin fiber. However, using more detection channels will also increase the overall scan time, as more lines on the camera will have to be read out. Likewise, the demands on the computer hardware for post-processing will be higher. In addition, dark and readout noises are increased which will at some point degrade the image quality for low signal yields.

We have performed measurements on a blend of microspheres consisting of polystyrene (PS, diameter 0.5 μm) and polymethylmethacrylate (PMMA, varying diameter 1–10 μm), which were air dried on a coverslip. The sampling distance was chosen to be 100 nm.

Confocal Raman images of this sample are shown in Fig. 5(a). The images represent sums over spectral bands which are specific for PS and PMMA.

 

Fig. 5. Measurements on a PS/PMMA mixture; first column, confocal Raman images without the pixel reassignment; second column, with the pixel reassignment. To highlight the resolution gain, the signal profiles through the bulk of the PS beads are shown in (b).

Download Full Size | PDF

The first column contains images of a standard confocal measurement. This imaging configuration was mimicked by just adding up the signals from the individual fibers without the described post-processing method, therefore ensuring exactly the same imaging conditions, i.e., collected intensity and matching resolution, as for our proposed method. The second column contains the corresponding images after the pixel reassignment. The improved resolution is evident: in the bulk area, the pixel reassignment allows discriminating of the individual PS beads, while the standard confocal image only shows an almost homogeneous signal. The signal profiles through the bulk in the two images are shown in Fig. 5(b). In addition to the improved resolution, the effect of a super concentration of light [13] is visible in the form of increased maximal pixel values in the pixel-reassigned image (red plot).

In view of photon shot noise, which is often the dominant noise source for modern detectors, the pixel reassignment represents a clear advantage. In our case, the signal gain of 2.7 leads to a SNR improvement of about 1.64. Even at the lowest signal levels, where camera-related noise such as read out and dark noise become dominant, the noise increase of, in our case, $\sqrt{7}$, is fully compensated for by the signal gain, such that the SNR ratio in the pixel-reassigned image is always at least comparable to that obtained with single-fiber collection at equal resolution.

In some experiments, we noticed lower noise levels in the reassigned images, even in dark zones with no signal. We attribute this effect to temporal averaging, which the pixel reassignment inherently provides by collecting signals from one specific spatial scan point at different points in time. This averaging effect might help mitigate undesired periodic fluctuations in the system of any sort (e.g., laser instabilities).

In summary, we have presented an easily implementable modification to fiber-coupled Raman microscopes that allows one to improve the lateral spatial resolution without sacrificing light. This is achieved by incorporating the pixel reassignment, a method originally developed for confocal fluorescence microscopy. We have realized the required pixelated detector by employing a round-to-linear fiber bundle in conjunction with a multi-line CCD-based spectrometer. A lenslet array in front of the fiber bundle ensures a high collection efficiency of close to 100%. In contrast to fluorescence microscopy, where pixel dwell times are typically on the order of microseconds, the scan speed in confocal Raman microscopy is usually much slower (an integration time of 0.2 s was used for all shown data) because it is limited by the strength of the relatively weak Raman signals. These slow scan speeds make our approach feasible by ensuring that the increased CCD readout time does not represent a “bottleneck” in the image acquisition process. We experimentally confirmed the improvement in the lateral resolution and the signal gain, respectively.