Cell Optical Density and Molecular Composition Revealed by Simultaneous Multimodal Label-Free Imaging

Abstract

We show how Raman imaging can be combined with independent but simultaneous phase measurements of unlabeled cells, with the resulting data providing information on how the light is retarded and/or scattered by molecules in the cell. We then show, for the first time to our knowledge, how the chemistry of the cell highlighted in the Raman information is related to the cell quantitative phase information revealed in digital holographic microscopy by quantifying how the two sets of spatial information are correlated. The results show that such a multimodal implementation is highly useful for the convenience of having video rate imaging of the cell during the entire Raman measurement time, allowing us to observe how the cell changes during Raman acquisition. More importantly, it also shows that the two sets of label-free data, which result from different scattering mechanisms, are complementary and can be used to interpret the composition and dynamics of the cell, where each mode supplies label-free information not available from the other mode.

Introduction

A main challenge in recent microscopy is to push toward gathering more information about the observed sample, while being subject to requirements such as acquisition speed and signal/noise. Going beyond the limitations has been achieved in a variety of ways, such as improving specificity through higher spatial resolution, enhancing sensitivity by employing more efficient dyes or more sensitive detectors, or increasing the amount of measured channels in multimodal approaches. These developments also have to be undertaken while respecting a number of constraints such as acquisition speeds due to the limited amount of photons that can be detected for a given recording scheme, or the photo-damage that could be induced in the sample by ramping up the excitation power.

Multimodal imaging in microscopy attempts to overcome limitations in a given mode by combining it with a complementary mode. Spectral selection of fluorophore excitation and emission profiles, separation of different nonlinear optical phenomena such as second or third harmonic generation (1), employing light properties such as polarization (2), or additional complementary contrast mechanisms can all be combined to attempt to optimize the amount of information available from a sample. In the case where one of the modes is based on microspectroscopy, however, such simultaneous approaches are difficult to implement due to the necessity of measuring a large spectral range, and the preclusion of using other chemicals in the sample (3). Although the combination of several spectroscopic measurements can provide invaluable information, it typically requires great care in sample manipulation and needs sequential measurements with subsequent image registration (3,4).

In particular, Raman spectroscopy (RS) is a valuable tool for observing biological samples, due to its nondestructive nature and the weak scattering of water. Its imaging and spectroscopic information can be used at a cell imaging level, such as for the detection of metal complexes (5) within live cells, the identification of aging cells (6), observation of molecular dynamics during apoptosis (7), or the characterization of undifferentiated cells (8). It can also provide invaluable information at the tissue scale, such as diagnostics through spectral analysis for breast cancer (9), epithelium inflammation (10), discrimination of atherosclerotic plaques in blood vessels (11), or the functional imaging of full microorganisms (12).

One major limitation of RS, however, is its long acquisition times, required to collect enough scattered photons to build up a spectrum; this constraint becomes even more stringent when considering imaging systems, where two-dimensional scanning is required to construct the image.

Multimodal approaches can be employed to circumvent this limitation, by providing a fast imaging modality operating in parallel to the molecular specific RS mode. This approach has been employed for example on tissues, by coupling RS with optical coherence tomography, which can provide the general three-dimensional structure of the sample (13). We propose in this article a multimodal approach for microscopy, where we combined RS with digital holographic microscopy (DHM), an interferometric imaging method that can provide additional sample information by quantitative measurement of the phase shifts induced by the sample (14).

Several features of DHM make it an ideal candidate for combination with RS. It uses a noninvasive low-power laser illumination, and provides high frame rates, making it highly complementary to RS. It is also a label-free method, so that specimens can be observed without requiring specific protocols before measurements. Raman emission is usually wideband and unpolarized, which prevents most multimodal implementations because the signals from each mode cannot be separated. A combination of both DHM and Raman measurements, based on sequential measurements with the same light source has recently been reported (15). However, when employing two different light sources, the narrow bandwidth of DHM allows us to easily and simultaneously separate it from other signals (16). This approach makes it possible to measure both RS and DHM simultaneously, and allows us to monitor the sample in real time by DHM during RS acquisition.

Additionally, DHM provides quantitative information through the integrated phase shifts induced by the sample, which can be employed to derive biological indicators such as the dry mass (17) of cells. DHM was shown to provide valuable information in different biological applications, such as to monitor neuronal cell receptor activity during activation (18), follow cell differentiation (19) or life cycle (20) through morphological features, assess the inflammatory state of colon tissue (21), or determine cell viability (22), possibly during high-throughput screening assays (23).

The combination of both DHM and RS in a dual system makes it therefore possible to follow rapid morphological changes in one channel (DHM), while more specific information about the molecular vibrational state can be retrieved from the second, slower channel (RS). As both modalities enable measurement without any contrast agent, this approach keeps the advantage of label-free imaging, where the sample can be measured without any processing or insertion of labels. Interestingly, this also implies that while DHM provides information about the linear elastic scattering of a wave front passing through a specimen, RS measures the inelastic scattering of the specimen’s content at the wavelength of excitation. The combination of these two signals then provides a complete description of the linear response of the observed specimen, even though RS and DHM are usually regarded from very different standpoints in terms of their physical meaning.

We first describe in the Experimental Setup and Data Processing sections the experimental configuration we developed to enable simultaneous RS-DHM measurements, and detail the calibration procedures employed to enable comparison of the two types of images. In the first part of the results, we then present the capability of obtaining simultaneous multimodal measurements on HeLa cells, based on the methods given in the Materials and Methods section, and compare the images obtained with the two modalities. In a second stage, we refine the analysis of the Raman spectra, to derive a quantitative comparison between the spectral information and the quantitative phase one, with the aim of moving toward integrating both inelastic and elastic imaging modes.

Materials and Methods

Cell preparation

HeLa cells were cultured on a 3.5 cm quartz bottom dish (Fine Plus International, Kyoto, Japan) and immersed in Dulbecco’s Modified Eagle’s Medium (DMEM, Nacalai Tesque, Kyoto, Japan) with fetal bovine serum (10%, FBS, Nacalai Tesque), and incubated during 1-2 days at 37°C in a humidified atmosphere with 5% CO₂. Before observation, the culture medium was replaced by a phosphate buffered saline (PBS, Nacalai Tesque) buffer solution by washing the dish 3–4 times. Cells were then measured at room temperature.

Measurement procedure

All measurements were performed with the experimental setup described in the Experimental Setup section, and graphically represented in Fig.1. For phase measurement, cells were continuously exposed to the DHM laser with a power of ∼400 μW over the entire field of view provided by the microscope objective (MO). The charged-coupled device (CCD) detector then acquired images with a typical exposure time of 5 ms. Images were continuously recorded with a 1024 × 1024 resolution at a frame rate of 0.3 Hz.

Figure 1.

Optical setup enabling simultaneous Raman and digital holographic imaging.

During Raman measurements, cells were excited with a focused beam with a continuous wave power of 4.39 mW/μm². The spot was rapidly scanned vertically at 100 Hz to produce a line excitation during the 3 s integration time of the detector. The scanned region was typically chosen as 400 pixels in the vertical direction, with 200 horizontal lines recorded, thus corresponding to a total acquisition time of 10 min for one hyperspectral stack. All measurements were carried in a time window <1 h after cell preparation, a time frame suitable to maintain cell viability in the observation medium.

Data processing

The phase images were extracted from holograms through standard off-axis methods. The complex wave field can be retrieved from the hologram through Fourier filtering (24), and demodulated with a reference correction hologram (25) measured from an empty field of view in the dish. The complex field was then propagated into focus by digital means with an angular spectrum convolution kernel (26). The fact of being able to adjust the focus of the phase image after acquisition makes it possible to adjust the mechanical focus purely to optimize the excitation efficiency of the Raman channel. After propagation, the phase image was unwrapped to suppress phase discontinuities if required (27).

The Raman hyperspectral data set was first background subtracted (from an empty field of view), and baseline corrected to suppress remaining fluorescence. Baseline correction was performed by estimating the curve with the lowest probability of being larger than the signal at several data points with a low quantile value, which was then estimated on all samples with cubic spline interpolation. The spectra were then smoothed with a locally weighted scatterplot smoothing procedure, which is known to reduce noise while preserving spectral peaks (28).

Experimental Setup and Data Processing

Experimental setup

To enable the simultaneous observation of quantitative phase and Raman imaging, we exploit the fact that the laser-based DHM is spectrally limited, and can thus be separated from the Raman excitation and emission, provided that the DHM laser wavelength is set outside of the Raman range. A schematic showing our optical setup for the two combined modalities is shown in Fig. 1.

The Raman excitation is performed with a continuous wave laser at 532 nm (Verdi V-6, Coherent, Santa Clara, CA) and the light is tightly focused in the object plane with a microscope objective (MO, CFI Plan Apo IR 60XW, NA 1.27, water-immersion, Nikon, Tokyo, Japan). The Raman backscattered light passes through a short-pass dichroic mirror (DM) used to separate the DHM wavelength, and is separated from the Raman excitation beam with a long-pass DM, before being focused on the slit of a spectrometer with a relay optics (RO). The spectrometer (Shamrock, focal length 500 mm, 300 lp/mm, Andor Technology, Belfast, Northern Ireland) separates the Raman frequencies and images the spectral line on a low-noise CMOS camera (Orca-Flash 4.0, Hamamatsu Photonics, Hamamatsu, Japan). To perform imaging through scanning microscopy, a first galvano-mirror (GM1, Cambridge Technology, Bedford, MA) scans the beam in a direction parallel to the entrance slit of the spectrometer at a frequency of 100 Hz, much higher than the exposure time of the detector. This pseudo slit-scanning configuration allows one frame of the camera to capture a two-dimensional data set containing all spectra of one spatial line, to minimize Raman imaging time (29). A second GM (GM2), placed after the long-pass DM on a conjugate plane of GM1 with a set of lenses then displaces the beam in the direction perpendicular to the spectrometer slit between each frame acquisition.

The phase imaging part of the experimental setup consists of a Mach-Zehnder interferometer, in which the sample is observed in transmission. The light emitted by a laser diode (VCSEL-780, Thorlabs, Newton, NJ), has a wavelength $\left(\lambda \simeq 780\text{nm}\right)$ chosen to be longer than the longest Raman shift of interest. The beam is split into two beams by a beam splitter, and the object beam is spread by a beam expander and focused by a condenser to evenly illuminate the field of view. The scattered light emanating from the specimen is collected with the MO and separated from the Raman excitation/emission with the short-pass DM, while RO projects back the image of the specimen near the detector. The reference beam is recombined with a second beam splitter with a small angular offset compared to the object beam in an off-axis configuration (24). A CCD camera (TXG20, Baumer, Frauenfeld, Switzerland) then records the interference pattern between the two coherent waves.

The observation protocol used with this experimental setup and the data processing employed to extract the signals are described respectively in the Measurement procedure and Data processing sections.

Image registration between two modes

To enable comparison between the images obtained through RS and DHM, it is first required to adjust the two fields of view, as they were acquired with completely independent imaging modes. The two modes therefore may differ in field of view, alignment of the optical elements, crop/scaling factors induced by different sensor/pixel sizes of the two cameras, and differences in magnification of the relay optics.

In the employed configuration, DHM provides the larger field of view, so that the phase image was fitted onto the obtained Raman image by averaging in the C-H stretching region (2890–2960 cm⁻¹). The multimodal image registration was then performed by maximizing the cross correlation between the Raman image and a transformed phase image through scaling and rotation. The translation was then determined by locating the maximal value of the cross correlation. This procedure was iteratively improved through an unconstrained nonlinear optimization, implemented with the fminsearch function of MATLAB (The MathWorks, Natick, MA). After convergence, the phase image was transformed with the determined fitting factors and downsampled to match the pixel resolution of the Raman data set to ease comparison.

Independent component analysis

For multivariate analysis of the Raman data, we employed independent component analysis (ICA) (30), which is comparable to the well-known principal component analysis (PCA) method. PCA decomposes a Raman data set into uncorrelated vectors for easier understanding of the underlying signals, by projecting data onto a new basis that maximizes the variance on its projection vectors. ICA relies on a similar principle, where it maximizes the so-called non-Gaussianity of data on its projection axes. The main difference between the two approaches relies on the fact that while PCA projects data on an orthonormal basis, the vectors of ICA do not follow this constraint, thereby making it possible to find components that have spectral overlap, and more closely match the actual components in the sample rather than separable components.

In detail, it is required for ICA to employ an indicator of non-Gaussianity, which is classically chosen as the third-order moment, or kurtosis, defined as

$$\text{kurt}\left(\mathbf{y}\right)=E\left\{{\mathbf{y}}^4\right\}-3{\left(E\left\{{\mathbf{y}}^2\right\}\right)}^2,$$ (1)

where $E\{\cdot \}$ is the expectancy. More intuitively, we can consider that white noise is characterized by a normal distribution, so that ICA projects data on a vector basis in which noise is best rejected by selecting directions for which the distribution is the furthest from a normal distribution.

We performed ICA by employing the fastICA package for MATLAB, developed by A. Hyvärinen et al. (31). This implementation extracts the independent components (IC) through a fixed-point iteration approach.

We should note that some preprocessing steps are required before performing ICA, which relies on several hypotheses on the input data. In particular, the data must be mean-centered, which corresponds to subtracting the mean of the signal, i.e.,

$${\mathbf{x}}^{\prime }=\mathbf{x}-E\left\{\mathbf{x}\right\}.$$ (2)

Furthermore, the data must be prewhitened, to ensure that input vectors are uncorrelated. This corresponds to diagonalizing the correlation matrix ${\mathbf{C}}_{\mathbf{x}}=\mathbf{x}{\mathbf{x}}^T$, so that the data used as input for ICA $\tilde{\mathbf{x}}$ is

$$\tilde{\mathbf{x}}=\mathbf{E}{\mathbf{D}}^{-1/2}{\mathbf{E}}^T{\mathbf{x}}^{\prime }.$$ (3)

In practice, it is also usually helpful to reduce the dimensionality of $\tilde{\mathbf{x}}$ to shorten computation time and improve the stability of the output. This implies that Eq. 3 can be implemented through standard PCA, and the generation of $\tilde{\mathbf{x}}$ can be limited to a chosen amount of eigenvectors.

Results and Discussion

Multimodal imaging

We present here measurements recorded with the arrangement shown in Fig. 1, where we continuously recorded phase images during acquisition of Raman hyperspectral stacks. The measurements are performed on HeLa cells, a well-known adherent cell line, which is ideal for observations such as the ones described below, where global cell movements are minimal, making it easier to compare the intracellular dynamics between the two modes. The field of view in phase shown in Fig. 2 A corresponds to the HeLa cells at the beginning of the experiment, and can be compared with the Raman acquisitions (cf. Fig. 2, B and C), represented in the lipid bands (2850 cm⁻¹) and cytochrome c (1315 cm⁻¹). Because recording one Raman stack requires 10 min, data were recorded over the intervals 120–720 and 1056–1656 s. As the frame rates of the phase and Raman cameras were both 3 s, each phase image corresponds to the acquisition of one spectral line. This is represented in Movie S1 in the Supporting Material, where the cell morphology can be observed to change in phase during the Raman acquisitions.

Figure 2.

Simultaneous measurements of HeLa cells with (A) quantitative phase digital holography, and Raman images (B and C), shown here in the lipid (2850 cm⁻¹), and cytochrome c (1315 cm⁻¹) bands, taken ∼15 min apart. The Movie S1 also shows the temporal evolution during the course of recording.

The DHM image in Fig. 2 A shows the cell content in phase; with increasing intracellular content, the relative phase shift is increased, leading to higher brightness values within the cell cytosol. Denser regions can be observed at the center of cells, in particular the bottom-right cell, corresponding to the cell nucleus. One can also observe smaller structures in the cells, which are moving rapidly in the time-resolved phase measurements (cf. Movie S1).

The Raman image shows the main cell shapes by selecting a protein band in green (cytochrome c), with lipid aggregations in red. We can observe that some lipid structures are elongated in the vertical direction (see arrows in Fig. 2), in particular in the left cell. This feature shows the usefulness of the multimodal method: by comparison with the phase images over time, we can determine that these elongated lipid structures in fact result from motion blur, appearing aligned with the slow scanning axis (vertical) in the Raman images in Fig. 2, B and C. These dense structures are also observed to be rapidly moving in the phase images.

In practice, the DM spectrally separating the Raman and phase imaging modes possesses some leakage, so that the Raman excitation laser that irradiates the sample also impinges on the DHM CCD camera, but we can determine that there is no influence of the green excitation beam on the phase signal, despite the high power involved at the excitation stage. This is because the green laser is not coherent with the infrared laser diode employed for holographic recording, and corresponds only to an additional incoherent term, which is inherently filtered out during the phase reconstruction procedure. Similarly, we also note that the near-infrared laser diode light for DHM imaging does not interfere with the Raman acquisition. Tests showed that the diode laser leakage reaching the Raman detection camera is of the same order of magnitude as the Raman signal (data not shown), and is in any case located spectrally at a wavenumber of around 5975 cm⁻¹, outside the Raman signal range of interest.

We can observe some morphological changes between the two Raman measurements shown in Fig. 2, B and C, such as slight cell shrinkage and reorganization of lipids aggregation, which can be further understood by analysis of the phase images in the Movie S1. In another example, shown in the Movie S2, it is also possible to identify organelle movements in the cytoplasm during the Raman recording. Some of the observed cell shrinkage may be due to stress induced by the Raman laser excitation, which has absorption at 532 nm. This is corroborated by the fact that cells seen in phase but outside of the measurement area of the Raman mode seem to be less prone to shrinkage. The rapid phase channel enables the monitoring of these changes in real time to assess cell viability during measurement through morphological features, to ensure that cells are viable throughout and after the Raman measurement. It also makes it possible to account for motion when interpreting the Raman spectra, and provides additional insight about the cell dynamics in parallel with the spectral information.

Sources of image contrast in phase versus Raman modes

By performing simultaneous Raman and quantitative phase imaging, we can study in more detail the biophysical sources of respective contrasts in the Raman and phase image modes. For comparison, we first adjust the two fields of view to make them comparable with the image registration procedure described previously, yielding the results shown in Fig. 3, where both images now have the same sampling conditions.

Figure 3.

Comparison between phase images (A and D) adjusted with the previously described image registration procedure Raman images (B and E) shown in the lipid (2850 cm⁻¹) and cytochrome c (750 cm⁻¹) bands, and an average of the Raman spectra (C and F) in the C-H stretching region (2890–2960 cm⁻¹). The dashed cross in Fig. 3C represents the region from which data has been extracted to generate Fig. 4.

We selected two fields of view where cells showed different characteristics. The cell in Fig. 3, A–C, has a round shape, with dense lipid droplets in its cytoplasm and is seemingly in the interphase state, whereas the cells shown in Fig. 3, D–F, have a more triangular shape with an even distribution of lipids in the cytoplasm, and appear to be at the end of mitosis. This can be identified by the nuclei of the two cells being close together (low lipids regions in the Raman image, Fig. 3 E), while being gradually separated in the phase image, with denser regions within the nucleus. We can immediately note that the overall cell shape is similar between the phase (Fig. 3, A and D), and Raman images, in particular when looking at the Raman distribution from the C-H stretching region (Fig. 3, C and F).

Going deeper with the comparison, both methods are label-free, so that their contrast originates from light interaction with intracellular compounds. However, the specific contrast mechanisms are significantly different. DHM results from elastic scattering of photons, retarding the wave front, which is then observed as phase differences. Furthermore, it is a wide field microscopy technique, where the signal is generated by the propagation of the wave front through the specimen, without optical sectioning. The Raman images result instead from inelastic backscattering of photons. Because it is performed in point or line-scanning mode, with a pinhole or slit over the detector, a degree of optical sectioning is achieved. This implies that the signal retrieved from the Raman mode may not account for the whole depth measured by DHM. In the case of our experimental parameters, the depth of the excited region is in the order of 1 μm, and therefore accounts for a significant amount of the depth content in the case of HeLa cells that are adherent. The two signals may however become less comparable for significantly thicker specimens.

As observed in Fig. 3, despite these differences in physical contrast mechanisms and optical sectioning capability, we do observe enough overlap between the two modes to enable image registration and to treat the two sets of information as complementary. The comparison between phase (Fig. 3, A and D), and the C-H stretching region of the Raman signal (Fig. 3, C and F), present similar intensities in the cell cytosol, with a clear boundary around the nucleus, and local aggregations. These similarities are however distinguishable in Fig. 3 only with one particular spectral region of the Raman data; the choice of which Raman region to compare to the phase is not trivial. We therefore assess the spatial similarity between the two modes across the entire Raman spectrum, by computing for each measured wavenumber the Pearson cross correlation between the Raman and phase images, defined as

$${\rho }_{\phi ,I_R\left(\omega \right)}=\frac{E\left[\left(\phi -{\mu }_{\phi }\right)\left(I_R\left(\omega \right)-{\mu }_{I_R\left(\omega \right)}\right)\right]}{{\sigma }_{\phi }{\sigma }_{I_R\left(\omega \right)}},$$ (4)

where φ and $I_R\left(\omega \right)$ are respectively the phase and the Raman image at a given wavenumber, stored in one-dimensional vectors; μ and σ are the mean value and standard deviation over the spatial distribution, respectively. The result of the cross correlation computation for the cell in Fig. 3, A–C, is shown in Fig. 4, and compared with a spectrum taken on a $5\times 5$ square region in the cell cytosol, at a location indicated by the dashed cross in Fig. 3 C.

Figure 4.

Comparison of how phase and Raman spatial information are correlated across the entire spectrum. The Pearson cross correlation between the phase and Raman images for each wavenumber is compared with a Raman spectrum from the cytoplasm of the cell shown in Fig. 3, A–C. The Raman spectrum has been normalized to ease the comparison with the correlation values.

From the result, we can see that the cross correlation follows the shape of the spectrum (in this case chosen from the cell cytoplasm). This means that regardless of the wavenumber, the overall spatial distributions are correlated between Raman and phase. It additionally shows that spatial distributions from regions of higher signal in the spectrum are more highly correlated with the phase distribution. This would be explained by the fact that the presence of molecules in the cell produces a signal in both modes. In particular, regions with large spectral peaks, such as the C-H stretching region, and the large peaks at 1450 and 1660 cm⁻¹, exhibit high correlation with the phase signals. Some smaller features are not present however, and the correlation appears to be far smoother than the spectrum.

This then leads us toward the possibility of further analysis of which molecules primarily contribute to the contrast of the phase imaging mode. This approach has meaning, not only for phase-Raman multimodal imaging, but for phase imaging in general such as DIC and phase contrast, since the phase information is independent of the measurement method used.

Independent component analysis

To derive more specific links between the chemical and phase information data sets, we employ the ICA method to decompose the hyperspectral data (30). In particular, it has been shown that ICA can provide a more specific decomposition, which is also easier to interpret thanks to the general absence of negative peaks (32,33).

Employing multivariate analysis makes it possible to generalize the study, and to derive more representative spectra of cells, we selected three additional measurements to the ones shown in Fig. 3, and extracted the locations that contain the cells in the five fields of view. The spectra at these locations were treated as described in the Data processing section, and then extracted to create one multicell data set consisting only of cellular Raman data. The data matrix was then mean-centered and whitened as described in the Independent component analysis section, and reduced in dimensionality to N′ to improve the stability of convergence of the algorithm, by selecting the eigenvectors representing the first 50% of the data variance, before performing ICA. In contrast to PCA, the resulting components are not ordered, so that it was necessary to manually determine the most relevant vectors. We selected the six most relevant ICs by ensuring they had the strongest spatial contrast (by inspection). This led to the vectors shown in Fig. 5, where the hyperspectral data of the cell shown in Fig. 3, A–C, is projected back onto the respective ICs. By comparison, employing PCA on the data set (not shown) imparts a requirement of orthonormality on the components and led typically to only two or three main components with identifiable spatial contrast, which is less than the number shown in Fig. 5.

Figure 5.

Selected IC vectors derived from five measurements, with the corresponding projections of the cell shown in Fig. 3, A–C. The wavenumber values discussed in the text are shown numerically in the respective plots.

It is possible to see that the various ICs project onto different sets of spatial information, such as IC 1, which spatially outlines the general shape of the cell with a projection vector resembling the global shape of a cell spectrum, with strong peaks in the CH bonds (1448, 2934 cm⁻¹), and peaks commonly attributed to α-helix structures (1324, 1657 cm⁻¹) (34), along with a significant contribution in the 700–1100 cm⁻¹ region. Interestingly, IC 1 presents very similar characteristics to spectra obtained through vector component analysis (VCA) on other cell types, also attributed to protein vibrations (35).

ICs 2 and 3 clearly select the lipids with CH₂ features at 1458 cm⁻¹ and distinctive peaks in the C-H stretching region. The presence of C=C vibrations (1650 cm⁻¹) also indicates unsaturated lipids (36). The two components present different spatial contrasts, as well as some spectral differences. In particular, the more pronounced CH₂ rocking band (IC 3, 1295 cm⁻¹) and the shifts in several bands may indicate different types of lipids (36,37). However, due to the spatial features with a directional preference for each IC and the fact that several bands are shifted, it is impossible to fully exclude contributions from optical aberrations.

The next ICs are then more difficult to interpret spectrally; ICs 4 and 5 correspond to the global cell shape, with IC 4 having an additional gradient within the cell region, and a vector that still possesses features similar to a Raman spectrum (C-H stretching or 1650 cm⁻¹ regions), but also some negative regions which are difficult to interpret. On the other hand, IC 5 resembles a spectrum with lipid bands (1440, 1650 cm⁻¹), along with some strong contributions in the low wavenumber region. Finally, IC 6 spatially selects the nucleus, and logically possesses a spectrum with less C-H stretching and some peaks in the 850–1050 cm⁻¹ region.

One has to note that as the ICA was performed from the data of several cells, the components are not necessarily the ones providing the least noisy projections, but the ones providing the most representative vectors for all considered cells. The projections of the other cells employed to derive the IC provide similar spatial contrasts as the ones shown in Fig. 5.

Phase components assignment

Having decomposed the Raman spectral data into several components with various spectral characteristics and spatial contrasts makes it possible to quantitatively estimate the respective contributions of the extracted spectral components to the quantitative phase images. To this end, we computed the linear combination of these components that most closely matches the phase signal, and thereby estimate the respective contribution of each Raman-active component to the phase. This can be performed by minimizing the following least-square problem

$$\beta \left(\mathit{\alpha }\right)=\sum _{x=0}^N\sum _{y=0}^M{\left(\phi \left(x,y\right)-\sum _{i=0}^P{\alpha }_i·{\text{IC}}_i\left(x,y\right)\right)}^2,$$ (5)

where $\phi \left(x,y\right)$ is the phase image, ${\text{IC}}_i\left(x,y\right)$ is the projection of the hyperspectral data onto the ith IC, N, M are the dimensions of the images, P is the amount of IC employed for fitting, and ${\alpha }_i$ are the coefficients of the linear combination. An example result of the linear combination of the components of Fig. 5 is shown in Fig. 6 A where the resulting values of the different ${\alpha }_i$ coefficients govern the reconstruction of the image shown in Fig. 6 B. It can easily be seen that the signal generated from the linear combination is closer to the phase image than any single one of the previously observed bands—lipids or cytochrome c (cf. Fig. 3)—or any of the single ICs (cf. Fig. 5). It has more constant contrast in the nucleus region compared to the C-H stretching region, and with a reduced contrast of the lipid aggregates, while still preserving the overall cell shape. This represents a main strength of our multimodal approach. We can determine which molecular components act cooperatively to form the phase contrast, and the results will be valid for any phase imaging method.

Figure 6.

(A) Linear combination coefficient values resulting from fitting the ICs of Fig. 5 onto the phase image of Fig. 3A, leading to an adjusted image (B).

These characteristics can be easily related to the coefficients values (cf. Fig. 6 A), where the IC 1 (cell proteins) is predominant, while the lipids (IC 2-3) are smaller. Similarly, the additional cell shape (IC 5) and the nucleus (IC 6) components possess relatively rather strong coefficients, as they are similar in strength to the lipid ones, although the ICs carry less energy. On the contrary, IC 4, which represents the cell shape but with a linear gradient within the cell appears to have a negligible contribution.

We then apply the same fitting procedure between the ICs of Fig. 5 and the respective phase images on a larger amount of cells (14 measurements, comprising in total 27 cells), to assess the robustness of the coefficients. The results, displayed in Fig. 7, show that the IC-based decomposition of the Raman data is robust and that the link between phase and molecular data remains very similar for all cells imaged, independently of the morphology of the cell, and indicate that the molecular components which contribute to the phase do not vary widely between different cells, as consistent with the assumption that each cell globally possesses similar intracellular concentrations.

Figure 7.

The Raman contributions to phase contrast are similar for all measured cells. Average value and standard deviation (error bars) of the contributions of ICs to match the respective phase image for n = 14 measurements, containing in total 27 cells. The assignments of spectral components as discussed in the text are denoted by different tones.

It is possible to see that all coefficient values are fairly reproducible throughout all measurements. ICs 1, 2, and 5 possess a smaller standard deviation relative to the others. IC 4 remains at a negligible level in all cases; interestingly, it also corresponds to the component whose projection vector is the most difficult to interpret as a spectrum (see Fig. 5). IC 3 and 6 possess consistent values over all measurements, but with relatively higher standard deviations. For IC 3, this higher variability may be explained by the fact that it accounts for lipids in conjunction with IC 2, whereas its spatial image is less contrasted than the droplets that can be identified in IC 2, so that the contribution of IC 3 to the phase signals varies more than in the other components.

It is interesting to consider the reproducibility of the link between the Raman and phase data for several measurements. The results of Fig. 7 show that the correspondence between the independent components and the phase distribution is relatively independent of the measurement conditions, despite the differences in physical interaction. Furthermore, as both signals (phase and spectral components) are originating from the intracellular content, it is likely that the correlation between these contributions are relatively independent from the cell type, as the measurements presented here are already the contributions from the whole cell body taken as an average. These results provide some degree of insight into the respective contribution of the different components related to various molecular structures to the phase signal, which is worth further pursuit. Although the typical view in phase imaging is that all intracellular content contributes to the measured phase shift, we can posit that some components have a stronger influence than others, and that these can be partially unraveled by a multimodal approach such as we have used here.

In terms of the physics of the two modes, phase retardation, measured in wide field, is linked to the refractive index, which originates from the permittivity tensor $\text{ϵ}$, a macroscopic quantity relating the electric field with the electric displacement in a material within the continuous model of Maxwell equations (38). On the other hand, the Raman measurement, recorded in point or line mode, reflects the inelastic scattering of the electron cloud in the probed molecules; the observed bands are related to the derivative of the polarizability tensor $\mathit{\alpha }$, which is a semiclassical representation of the dipole moment induced in the molecule by an external electric field (39). For simple molecules it is possible to relate the permittivity and the mean polarizability through the Lorentz-Lorenz relation (38). For the molecular distribution within a cellular body, however, such a relation is far from trivial. Most of the difficulty in relating both quantities lies in the fact that the refractive index is derived from a purely continuous model, even though several recent reports describe experiments probing refractive index at super-resolution (i.e., beyond macroscopic) scales in three-dimensional imaging (40,41).

We are then forced at this point in the comparison to take a qualitative view of the phase-Raman link for complex samples. For example, because large molecules are more likely to have a large permittivity, dense regions of such molecules such as proteins may have a stronger influence to the phase signal. The results of Fig. 7 appear to corroborate this statement, as ICs 1 and 5 are relatively strong, consisting of a general cell shape spatial contrast with a vector mainly composed of C-H stretching and peaks that can be commonly attributed to proteins, as discussed previously. On the contrary, the vectors attributed to lipids (ICs 2 and 3) have a strong contribution in the original Raman data, but contribute far less to the adjusted signal, showing that lipids have a smaller influence on the global phase signal. This is also supported by the spatial contrast in phase, where lipid droplets can be identified on the border of cells (where the rest of the intracellular content is low), but usually not at the center (see for example Fig. 2).

Finally, it appears that IC 6, attributed to the nucleus, has a significant phase contribution. This is consistent with the fact that DNA is known to have a small Raman cross section, even though it is composed of rather large molecules. This is evident in the nuclei of cells, where aggregates can often be observed in phase, and difficult to discern in the Raman mode. This can explain the strong contribution of this IC, and show that phase can also result from a significant contribution by DNA material. Additionally, the significant differences in the spatial distribution of this component may also explain the larger variability of IC 6 in its phase contribution (see Fig. 7), as the phase shifts in this region can either be rather evenly distributed or present aggregates seemingly being nucleoli.

In summary, the differences in physical contrast mechanism between Raman and phase make them ideally suited for multimodal use to highlight different types of molecules in the cell. For simpler molecular distributions where the relevant tensors are known, it could be extended to quantitative mapping of scattering mechanisms.

Even with the qualitative limitation due to the complexity of the sample, the combined multimodal label-free approach presented in this article makes it possible to derive reliable empirical relations between the molecular content, a nano- to microscopic quantity, and the phase shift, which is linearly related to the refractive index. This approach allows the determination of the molecular sources of phase contrast for DHM, DIC, and other phase imaging methods.

Conclusion

We developed a multimodal microscope that enables the simultaneous measurement of RS through a laser-scanning scheme and quantitative phase through DHM, and demonstrated its applicability on live cells. This RS-DHM multimodal capability makes it possible to retrieve the features of both approaches, with on one hand the molecular specificity of RS, and on the other the recording speed of phase imaging. With signal separation in the spectral domain, the two signals can be recorded independently, making it possible to follow the evolution of the sample during the recording of RS, retrieving molecular specificity, and to monitor the cell morphology in real time. As both methods are label-free, the sample preparation is kept minimal, and cells can be observed without requiring any specific protocol, apart from immersion in a proper observation medium such as saline or cell culture media.

In addition, both measurements provide information about the linear response of light interacting with the sample. While RS provides molecular bands based on the inelastic scattering of an excitation beam, the signal provided by DHM is the complex measurement—amplitude and phase—of the elastic response of the light scattered by the sample. Although it is not currently possible to quantitatively relate these quantities, both are ultimately originating from the polarizability tensor of each probed molecule. We compared these two signals by cross correlation, showing that on first inspection, they are sufficiently similar to register the two data sets, and we found that the main influence on their cross correlation is the signal/noise ratio of the Raman response. We then employed multivariate analysis to decompose the RS signal into more specific components, applying ICA, which was shown to provide more specific projection vectors compared to the more widely used principal component analysis, thanks to the absence of orthogonality constraints. These ICA vectors could then be employed to quantitatively relate the contribution of molecular species such as lipids and proteins to the phase shifts induced by the cells, providing an insight into the molecular source of phase contrast imaging modes in a cell. These relations were shown to be reproducible throughout different measurements even when containing different cell morphologies, and indicate that the quantitative phase signal originating from cells is primarily induced by molecules distributed throughout the whole cytoplasm such as proteins, and by dense regions of DNA.