2.1.

Lensfree Imaging Sample Holder

The sample, typically a 5-μl droplet containing bacteria, is deposited and left to evaporate for 2 min on a quartz cover slip (TedPella Inc., Redding, California, $19\times 19\times 0.5\mathrm{mm}$) that was placed on an 8-bit $2592\times 1944$ CMOS sensor (MT9P031, Aptina Imaging, San José, California). This configuration [illustrated in Fig. 1(b)] implements the lensfree on-chip technique reported in Allier et al.¹² Briefly and as illustrated in Fig. 1(a), the image formed in transmission onto the sensor results from the interference between the light coming directly from the illumination source (here a laser beam) and the light scattered by the bacterial cell(s).

Fig. 1

Lensfree imaging module. (a) Schematic illustrating the principle of lensfree image formation. (b) Picture of the developed module used as the sample holder.

This technique succeeded in monitoring and counting cells, single bacteria, or viruses using a light-emitting diode as the illumination beam and a thin wetting film enabling enhancement of the micron-sized particles by creating microlenses-like liquid films on top of them.^12–14 In this study, no wetting film is used since it would be detected by Raman spectroscopy rather than the specific signal arising from the bacteria. Yet, the small aggregates obtained after evaporation are efficiently detected. We routinely detected aggregates composed of as little as 5 to 10 individuals. Depending on the laser spot size impinging the sample, the FOV can be varied from ${24\mathrm{mm}}^2$ to $100{\mu \text{m}}^2$. More precisely, these large and small FOVs are obtained when the laser spot size is about the sensor size and when it is comparable to the size of the bacterium to probe (1 μm), respectively. In this latter configuration, the image formed onto the CMOS sensor is due to forward elastic scattering only and, thus, reveals both the laser spot and the bacterium patterns (Fig. 2). The operator is then able to accurately monitor the alignment of the probe onto the sample.

Fig. 2

Lensfree images of single B. cereus illustrating lateral alignment process. On the left, laser probe is correctly aligned onto the bacteria. On the right image, the laser probe is $\sim 0.8\mu$m right misaligned. Central cross marks the position of bacteria.

In sum, the entire droplet as well as a zoomed view of a single bacterium can be easily observed and accurate lateral alignment of the laser probe is possible, thanks to this so-called lensfree (or lensless) based scheme. Moreover, a forward scattering pattern, the so-called lensfree image, can be collected for each probed bacterium in order to extract its morphological characteristics. In practice, the spot size is easily adjusted by translating the sample along the laser beam using a vertical translation stage mounted on the optical bench, and the XY laser probe alignment is achieved using a double translation stage (PI micos VT-75) mounted below the lensfree module, as illustrated in Fig. 3 and described in Strola et al.^15,16

Fig. 3

(a) Schematic of the integration of lensfree imaging in the Raman optical bench. (b) Illustration of the four-steps operation flow: (1) using a high Z value, thus large spot size, the entire droplet is observed with diffraction patterns corresponding to bacteria; (2) using intermediate Z value enables the operator to choose the bacteria of interest and align in XY the laser probe onto it; (3) using a Z value 70 microns above the Raman focal position, the interference pattern can be collected by the lensfree imaging module; (4) Raman focal position is found when the diffraction pattern blurred due to the laser spot size being smaller than the bacteria.

2.2.

Setup

Figure 4 shows schematics of the setup [Fig. 4(a) and a picture [Fig. 4(b)] of the lensfree imaging based Raman microspectrometer.¹²

Fig. 4

(a) Schematic of the optical architecture. (b) Picture of the apparatus consisting of (1) 532-nm continuous wave TEM00 laser head, (2) optical density (0.3), (3) razor-edge 45 deg of incidence filter, (4) $100\times$, $\mathrm{NA}=0.8$ microscope objective, (5) lensfree imaging module on which is placed the sample, (6) two notch filters and $f=50\mathrm{mm}$ plane-concave focusing lens, (7) optical fiber 105 μm $\mathrm{NA}=0.22$, (8) Tornado Spectral Systems prototype spectrometer, (9) vertical motorized translation stage, (10) and (11) translation stage in the plane of the optical table, XY.

A 532-nm laser (Spectra Physics Excelsior 532-50-CDRH, Santa Clara, California) is both the Raman excitation light and the alignment light. By contrast, most Raman microscopes integrate separated light sources and paths, thus increasing the complexity and the risk of misalignments. The 50-mW optical output power is reduced by means of an optical density down to few nanowatts during the alignment step and to 34 mW for the Raman spectra collection. A razor-edge filter (Semrock LPD01-532RS-25, Rochester, New York) steers the laser beam at 45 deg into a $100\times$ microscope objective (Olympus LMPLFLN, Japan) that illuminates the sample from above. The $100\times$ magnification combined with the very high laser beam quality ($M^2=1.02$) enables a spot size of $<1\mu \text{m}$ in diameter at the beam waist together with a good wave front quality, an important specificity for large field lensfree imaging.

When the sample is in the beam waist position [Fig. 1(a)], Raman scattered light is generated and collected by the microscope objective, transmitted through the razor-edge filter, filtered from elastic scattering by two notch filters (Semrock NF03-532E), and, finally, focused into the spectrometer optical fiber (Thorlabs M18L01, 0.22 NA, Newton, New Jersey) using a 50-mm achromatic lens. The system, therefore, operates in a confocal configuration, similar to a traditional Raman microscope. The optical fiber tip (105 μm) acts as the pinhole and yields an axial resolution of 2 microns. The spectrometer, APEX-532, is a custom-built unit consisting of the best features from Tornado Spectral Systems’ HyperFlux 532 spectrometer and Ocean Optics QE65000 detector (Hamamatsu, Japan, thermoelectric cooled). The HTVS technology allows this unit to generate both broadband (spectral range from $-4000$ to $4000{\mathrm{cm}}^{-1}$) and high-resolution (${7\mathrm{cm}}^{-1}$) spectra while enhancing the optical throughput in the wavelength range of interest for bacteria analysis range (500 to $1500{\mathrm{cm}}^{-1}$). This patented technology modifies the shape of the beam in the spectrometer with $>90\%$ total optical throughput and differs from conventional spectrumeters that use a narrow entrance slit to achieve higher resolution at the cost of throughput. HTVS technology alleviates the classical trade-off between resolution and throughput in a dispersive spectrometer: the light beam is expanded while the total flux is preserved, allowing for an improvement of the performance.¹⁷ As an example, Fig. 5 shows two spectra of the same polystyrene sample (1 mm thickness) acquired with a 1-s acquisition time on our platform and on the Horiba (Japan) LabRAM Aramis commercial system, respectively. Acquisition parameters were fixed to compare the two setups. The laser power was set to 34 mW in both systems and the spot size to 1 μm. As a side note, one may observe small lateral shifts originating from the difference in resolution as well as a slightly different calibration of the two instruments. We point out that this discrepancy does not affect later classification performance. We note that our setup performs better in terms of net intensity along the spectral range from 600 to $1800{\mathrm{cm}}^{-1}$. The SNR varies along the spectra depending on the light throughput curve that is different for the two acquisition systems. We calculated the SNR for the polystyrene peaks at 1001, 1193, and $1594{\mathrm{cm}}^{-1}$ to give a representative idea. Values are summarized in Table 1.

Fig. 5

In the spectral range from 600 to ${1800\mathrm{cm}}^{-1}$, the setup used in this study (dotted line) shows a better net intensity and signal-to-noise ratio (SNR) with respect to a classical Raman microspectrometer, in this case Horiba LabRAM Aramis (continuous line). We used 34-mW laser power at the polystyrene sample (1 mm thickness), with a spot size of 1 μm, in both measurements for 1-s acquisition time.

Table 1

Signal-to-noise ratio (SNR) values for the different spectrometers compared at selected peaks of polystyrene sample.

APEX-532 has been calibrated with Hg lamp calibration and polystyrene has been used as a daily reference sample.

Although the performance in SNR is in favor of the APEX-532 on very Raman-active samples, such as polystyrene, the situation is different for the very weak bacteria signals. Figure 6 shows representative bacteria spectra acquired at 30 s [Fig. 6(a)] and 60 s [Fig. 6(b)] on our instrument and the Aramis, respectively. SNR values are also included. We again note the distinct throughputs of the two instruments. Our instrument enhances the range between 600 and $1800{\mathrm{cm}}^{-1}$, which results in a lower peak height at ${2900\mathrm{cm}}^{-1}$ compared to the Aramis, which has a flatter response. In terms of SNR, the Aramis performs better at both 30 and 60 s integration. This was expected since, among others, the Aramis uses a detector cooled at $-60°\mathrm{C}$ compared to $-15°\mathrm{C}$ for the APEX-532. When increasing the acquisition time from 30 to 60 s, the measured SNR improves from 2.17 to 2.99 for the APEX-532, while it increases from 5.73 to 9.82 for the Aramis system. The lower relative increase in SNR for the APEX is explained by the fact that the instrument was optimized for short acquisition times and weak Raman signals.¹⁷ This is also in agreement with our choice of using APEX-532 in this application. The price paid in SNR for decreasing the integration time to 10 s is better offset by the gain in acquisition time with APEX than with Aramis.

Fig. 6

B. subtilis spectra and SNR values for acquisition time of 30 s (a) and 60 s (b) with our instrument (Bacram, simple line), and with the Aramis setup (in bold), respectively.

All these optical components, except for the spectrometer unit, are mounted on a single vertical translation stage (PI micos VT-75) with a resolution better than 0.4 μm, which allows an accurate adjustment of the focal position. Its large course (50 mm) also enables the adjustment of the illumination spot size onto the sample according to the lensfree based method previously described.

The translation stages, spectrometer, and CMOS sensor are controlled via a program developed under the software Labview® (version 2011). This program is a useful interface that allows full control of the setup. Alignment protocol takes $<1\min$ from the moment the droplet has been evaporated until the Raman acquisition starts. The Z position provided by the lensfree based scheme is $\sim 1$ to 2 μm off the correct Raman focal position, which is found by monitoring the appearance of the C-H band at $2925{\mathrm{cm}}^{-1}$ in the Raman spectra using minute translation steps (0.4 μm). An exposure time of 1 s is sufficient to obtain a Raman signal and guarantees a rapid alignment. Interference pattern collection is very fast, in the millisecond range, and the Raman exposure time has been decreased to 10 s, thanks to the high spectrometer throughput. Typically, it takes 25 min to collect 30 spectra from 30 different single bacteria in a single droplet.

Scattering patterns and Raman spectra are analyzed off line using MATLAB® (R2013a) and RStudio programs, respectively. This paper focuses on the Raman scattering analysis (preprocessing and classification techniques) to demonstrate the scheme’s feasibility to collect Raman comprehensive spectra and to identify bacteria. Scattering patterns analysis will be reported in a future paper.

Last, a direct imaging path was added to the system for validation purposes. The direct image was used (1) to facilitate system alignment, (2) to confirm that the particles detected in the lensless image were actual bacteria, and (3) to evaluate the detection limit in the lensless image. The detection limit was found to be small bacteria aggregates composed typically of 5 to 10 individuals. Direct imaging was implemented using a mirror in the Raman path. Light from a white illuminator is reflected by the mirror back to a CMOS sensor (Thorlabs, Newton, New Jersey) equipped with a Navitar objective. The magnification of this direct imaging path was set to 40. An illustration of the full optical path is presented in Fig. 4(a).

2.3.

Biological Protocol and Sample Preparation

Microorganisms E. coli (ATCC 9637), B. subtilis (ATCC 23857), S. epidermidis (ATCC 14990), B. cereus (ATCC 10702), B. thuringiensis (ATCC 33679), M. luteus (ATCC 4698), and S. marcescens (ATCC 27137) were purchased from American Type Culture Collection (Manassas, Virginia). B. subtilis, B. cereus, B. thuringiensis, M. luteus, and S. marcescens were grown overnight in Trypcase Soja Broth (Fluka 22092) at 30°C for Bacillaceae family and M. luteus and 26°C for S. marcescens. E. coli and S. epidermidis were grown overnight in Luria Broth (Sigma-Aldrich L2542, St. Louis, Missouri) at 37°C.

Each strain was cultivated in one broth culture in a liquid medium overnight (16 h). At the end of this culture step, all the bacteria have reached the stationary phase. Bacteria were then washed twice with Milli-Q ultrapure water by centrifugation (3500 rpm for 2 min) in order to ensure the complete removal of the medium. The pellet is then resuspended in 100 μL of Milli-Q water and absorbance measurements were made at $\lambda =580\mathrm{nm}$ (photospectrometer Uvicon923—BioTek Kontron, Winooski, Vermont). A stock solution with an optical density of 1 was then prepared, and Raman experiments were performed with a 1/100 dilution in MilliQ water of the stock solution.

The bacteria solution is then immediately processed with Raman acquisition to guarantee the biological homogeneity of the analysis during the stationary growth phase. An amount of 5 μL for each bacteria solution is pipetted on top of a quartz coverslip (TedPella Inc., Redding, California, $19\mathrm{mm}\times 19\mathrm{mm}$ and 0.5 mm thickness) previously rinsed with ethanol solution at 70% (Sigma-Aldrich, St. Louis, Missouri) and dried with nitrogen. This protocol ensured that all bacteria are in the same growth phase, independent of the strain prior to the Raman measurement. Before starting the measurements, we let the liquid drop containing bacteria evaporate to create an investigation region of a few millimeters in diameter. After each analysis, the quartz coverslip is carefully cleaned in an ultrasonic bath (Novatec, Baltimore, Maryland) for 10 min.

For each strain, we collected Raman spectra over 30 different single-cell bacteria at the stationary growth phase. This protocol allows us to minimize the biological variability of the various metabolism steps at the different growth phases. Each Raman acquisition has been performed with an integration time of 10 s. The background signal of the quartz coverslip, to be subtracted from Raman spectra of bacteria during data processing, is acquired at the same integration time at five random surface points before the deposition of the bacteria solution drop. Spectra were cropped to spectral regions of interest (ROIs) ranging from 650 to ${1800\mathrm{cm}}^{-1}$ and 2600 to ${3200\mathrm{cm}}^{-1}$, which cover the biochemical specific peaks of bacteria.^18,19

2.4.

Raman Spectra Analysis

Data analysis (spectra preprocessing, calculation of indicators, and classification) was performed using the R software environment, with existing functions or routines specifically developed for this use.

Preprocessing was applied to prepare the Raman spectra for the classification algorithm. The first step of spectrum preprocessing consists of cosmic spikes removal using the method proposed in Espagnon et al.²⁰ Then, several treatments have been considered for the input data to the classification method. From the simplest to the most complex, we tested the raw spectrum, smoothed and normalized, the first derivative, and the normalized net spectrum after background subtraction.

Smoothed signal and first derivative are calculated by Savitzky-Golay polynomials filters²¹ (degree 4, on 9 points). As the distance between two channels is ${7\mathrm{cm}}^{-1}$ in the ROI, a filter width of 9 points corresponds to ${63\mathrm{cm}}^{-1}$. This is to be compared to a full width at half maximum equal to ${57\mathrm{cm}}^{-1}$ for the quartz peak at ${800\mathrm{cm}}^{-1}$. The aim is to reduce the noise in the signal without peak distortion and loss of intensity.

Background subtraction is more complex and includes several steps. The background is composed of signals from the quartz substrate and the sample autofluorescence. For estimating the quartz signal, we consider the mean quartz spectrum, calculated using several quartz spectra acquired on the same coverslip at the same date as the bacterium spectrum. We then fit the mean quartz spectrum to each bacteria spectrum on the large peak spreading from 200 to ${650\mathrm{cm}}^{-1}$, specific to quartz [Fig. 7(a)]. An approaching method is presented in Beier et al.,²² with a rather different spectrum topology.

Fig. 7

Examples of E. coli spectrum: (a) raw spectrum and mean quartz spectrum after fitting on the main peak of quartz (both spectra were smoothed). The high contribution of the quartz in the bacterium spectrum can be observed. (b) Dotted line: spectrum obtained after subtracting the quartz contribution estimated by fitting. Solid line: spectrum after subtraction of quartz contribution and 600 iterations of the Clayton’s algorithm (specific net spectrum).

A constraint on the relative level of quartz signal, which has to be smaller than the bacteria spectrum on the region up to ${1700\mathrm{cm}}^{-1}$, is added to the fitting procedure. The fitted mean quartz spectrum is then subtracted from the bacteria spectrum. We deal with sample autofluorescence using the Clayton’s algorithm (also used in the sensitive nonlinear iterative peak-clipping algorithm²³), applied with a neighborhood window of three channels. This algorithm is iterative: when the number of iterations grows, the calculated background level drops. At the end of the process, the quartz contribution is extremely reduced and the bacteria peaks are, thus, emphasized, which facilitates their study. The resulting spectrum is called specific net spectrum [Fig. 7(b)].

Normalization is the last step before clipping the spectrum to the ROI. It is obtained by dividing the signal by its mean value on a chosen ROI. This enables us to have all spectra at the same scale, independent of factors that may vary between two spectra or two experiments (laser power, etc.).

For 100 spectra, preprocessing time is $\sim 8\mathrm{s}$ for data transfer and 7 s for cosmic spike removal. Background subtraction time depends on the number of iterations applied. For example, 600 iterations require 5 s, while 2500 require 30 s. These estimates are given for a PC equipped with an Intel Core i5 at 2.40 GHz and 8 GB of RAM. Processed spectra are stored in long-term memory for future reuse.

3.1.

Raman Spectra

Figure 8 shows typical images and spectra captured using our system. A blow-up of the two ROIs is displayed in Fig. 9.

Fig. 8

Validation of the lensfree ability to probe a single bacterial cell and collect scattering pattern and Raman spectrum. (a) Direct space white light image of the bacterial cell. (b) Backscattered image showing the laser probe is well aligned with the cell. (c) Forward scattered pattern collected using lensfree module. (d) Raman spectrum generated by the bacteria cell.

Fig. 9

Mean normalized net spectrum for E. coli and B. subtilis. Dotted lines correspond to three times the mean standard deviation of the mean value, divided by the square root of the number of spectra (154 and 273 for E. coli and B. subtilis, respectively). Raman bands detected are indicated and are found consistent with previous studies.²⁵

We display the average of processed Raman spectra acquired for B. subtilis and E. coli bacterial strains with a 10-s exposure time. The spectra consist of bands representing the cell contents: proteins, lipids, and nucleic acids. In particular, the peaks centered at 784, 1001, 1170, 1242, 1338, 1445, 1573, 1605, 1655, and $2925{\mathrm{cm}}^{-1}$ are detected for both strains.

Typical signatures of cell components are CH stretching vibrations and are visible at ${2925\mathrm{cm}}^{-1}$. The contribution from various proteins is related to a band centered at ${1655\mathrm{cm}}^{-1}$. In particular amide I, mainly associated with the $\mathrm{C}=\mathrm{O}$ stretching vibration and directly related to the backbone conformation, is detected in this band. Amide III, known as a very complex band dependent on the details of the force field, the nature of side chains, and hydrogen bonding, is revealed by the $1242\text{-}{\mathrm{cm}}^{-1}$ band.

DNA bands arise in the ${1573\text{-}\mathrm{cm}}^{-1}$ region via guanine and adenine nucleobases, but an overlapping with amide II contributions ($\mathrm{C}=\mathrm{C}$, N-H, and C-N stretching) can be reported.

${\mathrm{CH}}_2$ bending vibrations are assigned to the band at ${1445\mathrm{cm}}^{-1}$ and give contribution from various lipids.

Bands arising from amino acid side chains appear at 1605 and ${1001\mathrm{cm}}^{-1}$: the signals can be both assigned to phenylalanine. Tyrosine and phenylalanine give origin to the signal detected at ${1170\mathrm{cm}}^{-1}$. The ${1338\text{-}\mathrm{cm}}^{-1}$ region shows the fingerprint of the cytoplasm fraction detected via DNA vibration. The band centered at ${784\mathrm{cm}}^{-1}$ is covered by cytosine stretching vibrations as part of the DNA contribution. The peak assignment, based on previous works,^15,26 is summarized in Table 2.

Table 2

Assignment of Raman bands detected.

3.2.

Database Description

We conducted an acquisition campaign in order to populate a reference database composed of the seven strains of bacteria described in Sec. 2.3. Spectra were measured for four consecutive days, with 120 spectra per day (four strains of bacteria, 30 spectra each). In addition, spectra from other days were added to the database to reach a final count of 1205. Each spectrum was collected using an integration time of 10 s.

SDM and SNR for each strain were calculated (Table 3). SDM values are rather high, from 0.6 for S. epidermidis to values $\gg 1$, with 1.3 for E. coli and 1.7 for S. marcescens, mainly due to the low acquisition time. We find the same trend with SNR. Mean SNR ranges from 3 for the noisiest series (S. marcescens and E. coli) to 6 for the best series (M. luteus and S. epidermidis).

Table 3

Standard deviation of the means (SDM) and SNR values for the different bacteria investigated.

The dendrograms, PCA, and mean spectra, plotted for each strain, brought out some aberrant spectra, probably due to a contamination of the quartz substrate. That may explain the higher value of SDM obtained for M. luteus (0.8), compared to S. epidermidis, although SNR are similar.

3.3.

Preprocessing Choice

In this section, we investigate the various preprocessing options described above, namely, smoothed spectrum, first derivative, and net spectrum. We also discuss the choice of the ROI and number of Clayton iterations for computing the net spectrum. We use the classification score (calculated by cross-validation) as the figure of merit.

We first apply a fully randomized cross-validation (first scenario mentioned in Sec. 2.4.2). Figure 10 shows the classification score obtained with the specific net spectrum as a function of Clayton iterations number and ROI. As expected, the use of two ROIs rather than a single ROI yields better scores. The score plateaus around 84% after 3000 iterations. We, therefore, fix the number of iterations to 4000 for the net spectrum in the sequel. We now compare the performance of the various preprocessing options. For now, the net spectrum is computed using a quartz signal acquired on the same slide on the same date as the data. Results are summarized in Table 4. The first derivative method (74.7% of success) is outperformed by both smoothed spectrum (84.5%) and net spectrum (86.5%), with a slight advantage overall for the net spectrum.

Fig. 10

Classification rate according to the iteration number in the Clayton’s algorithm, for each region of interest (twofold cross-validation).

Table 4

Mean classification rates and standard deviation according to the type of preprocessing (fully randomized 10-fold cross-validation).

Then, we evaluate the preprocessing performance using the date-based cross-validation described in Sec. 2.4.2. Two methods for computing the net spectrum are now considered. First, the subtracted quartz signal was acquired on the same date. Second, the subtracted quartz is the mean of all quartz signals across all experiments. In this setting, we find that the net spectra calculated with the mean quartz and smoothed spectra have comparable performances (79.8% and 80.9% success, respectively, on the Bacillus family) and outperform the net spectra calculated with the quartz signal from the same date (75.9% success). Although these results do not indicate a significant advantage for quartz and background subtraction over a simple smoothing, it is noteworthy that no significant sample autofluorescence was observed in this dataset. The advantage of nonspecific signal subtracting would be more remarkable otherwise. With this in mind, we chose the net spectrum as the preprocessing method for the results presented in this work (with a quartz signal acquired on the same date).