A Raman-based endoscopic strategy for multiplexed molecular imaging

SERS Raman Nanoparticles.

The experiments performed in this study used SERS nanotags (Cabot Security Systems, formerly Oxonica Materials, Inc.) (19, 20). These Raman nanoparticles consist of a unique Raman-active layer adsorbed onto a 60-nm gold core and then coated with silica, resulting in an effective diameter of ∼120 nm. To demonstrate multiplexing, we used 10 different SERS nanoparticle flavors, each consisting of a unique Raman-active layer and a unique spectral fingerprint (SI Appendix, Fig. 1). We have described the molecular composition of these Raman-active layers previously (21). After excitation with a 785-nm laser, these SERS nanoparticles display a distinct spectrum based on the interaction of the incident light with the unique Raman-active layer adsorbed onto the gold core. The gold core acts as a surface field enhancer and can increase the effective Raman scattering efficiency by several orders of magnitude, which allows for ultrasensitive detection and makes it ideal for early detection diagnostic screening over large fields of view (14). Because the penetration of light is severely limited in tissues, we intend to integrate Raman imaging into an endoscopic instrument to access epithelial surfaces at clinically relevant depths within the body.

Design and Fabrication of Raman Endoscope.

The Raman endoscopic device uses a flexible fiber bundle with a centrally located single-mode fiber for illumination at 785 nm and an additional 36 surrounding multimode fibers for light collection. A singlet lens is used for dual purposes: to collimate the outgoing illumination beam and to collect the incoming SERS signals onto the multimode fibers. This collimating lens (6-mm focal length, 4-mm diameter) produces an illumination spot of ∼1 mm in diameter (Fig. 1A; more details are provided in SI Appendix, Collimating Beam). A unique feature of our Raman endoscope, as opposed to others already used clinically, is that ours is not a contact probe and has been specifically designed for efficient use, and signal uniformity, over a wide range of working distances, ranging from 1 to 10 mm, that are relevant for colonoscopy. This is necessary to accommodate user variability and imperfect centering during endoscopy, as well as the nonuniform surface topology of human tissue (SI Appendix, Working Distance). The circular honeycomb pattern of multimode fibers has been rearranged into a linear array at the proximal end of the fiber bundle, which is coupled into a spectrometer (RamanSpec RS-780; BaySpec) by aligning the linear array with the entrance slit (Fig. 1B, Inset). The spectrometer disperses the light collected by each multimode fiber, and the vertical slit is imaged onto a cooled, deep-depletion CCD array (DU920P-BR-DD; Andor Technologies) (Fig. 1C). The CCD then performs full vertical binning (FVB) of the sensor array to sum the spectral intensity of the collected light from all the collection fibers at each wavelength. Once an FVB acquisition is obtained for a given exposure time, the acquired signal is decomposed, using a direct classical least squares (DCLS) algorithm (Methods, DCLS-Weighting Factors), into weighting factors that quantify the intensities of the Raman-scattered signals as fractions of previously measured reference spectra.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 1.

Raman endoscope design and setup. (A) Schematic of Raman endoscope designed to be inserted through the accessory channel of a clinical endoscope with a 6-mm instrument channel. The Raman endoscope is composed of a single-mode illumination fiber that is surrounded by a bundle of 36 multimode collection fibers, totaling a diameter of 1.8 mm. The excitation laser light is collimated by a lens to emit an illumination spot size of ∼1.2 mm. (B) Photograph depicts the final fabricated Raman endoscope to be used for clinical studies. (Lower) Enlarged digital photograph of the endoscope head (Left), a magnified photograph of the fiber bundle (Center), and a magnified photograph of the back end of the device (Right) show a linear array of the 36 collection fibers that are specially aligned to fit into a spectrometer. (C) Schematic of the entire device setup starting with the 785-nm laser whose output is controlled by a shutter driven by a computer-driven shutter controller. The laser is then passed through a notch filter, which ensures a narrow 785-nm bandwidth, is guided through a series of mirrors, and is refocused to a single-mode fiber to illuminate a sample. The light collected by the multimode fibers is dispersed by wavelengths onto a CCD via a spectrometer.

The Raman endoscopic device designed for colon cancer detection has an external diameter of 5 mm (Fig. 1A) and is capable of being deployed through the instrument channel (6-mm internal diameter) of a clinical endoscope (GIF-XT30; Olympus). A major design challenge was to ensure that the Raman endoscope device fits inside the accessory channel of a conventional endoscope without damaging the optics. Most conventional endoscopes have an angular bend at the opening of the accessory working channel that requires the rigid length of the accessory to be short; it was imperative for the Raman endoscope to meet this particular design restriction (SI Appendix, Movie S1).

The current setup consists of a 785-nm laser that passes through a notch filter, two angularly adjustable mirrors, an electromechanical shutter, and a focusing lens to couple the laser (SI Appendix, Laser Coupling) into the single-mode illumination fiber (Fig. 1C and SI Appendix, Fig. 2). The electromechanical shutter (P/N:SH05; ThorLabs) is activated by a shutter control box (P/N:SC10; ThorLabs). The shutter is programmed only to open when the shutter controller receives a trigger signal from the CCD to acquire a signal. It is important that the shutter defaults to the closed position so that the exposure time of the laser to the tissue surface, for a given power and spot size, remains under the American National Standards Institute (ANSI) MPE (SI Appendix, ANSI MPE and Table 1). The notch filter is placed between the continuous-wave laser source and the single-mode fiber to eliminate broadband noise, including amplified spontaneous emission, from the laser. The maximum achievable power output from the distal end of the device is 42 mW (SI Appendix, Laser Characterization).

Our custom spectrometer was designed to have an N.A. of 0.28 to match our fiber bundle and CCD array requirements (SI Appendix, Spectrometer). The spectrometer has a wavelength range of 830–930 nm and includes an 840-nm long-pass filter to remove scattered incident laser light. The spectral resolution for the spectrometer system is ∼1 nm and resolves Raman transitions from 700 to 2,000 cm⁻¹. The spectral resolution was determined using a low-pressure mercury discharge lamp, which has distinct emission lines within the wavelength range that our spectrometer is able to detect.

Characterization.

The power output of the Raman endoscope was first monitored to determine its stability. The device’s power output was shown to be stable over varying working distances of up to 25 mm away from the sample surface; this was demonstrated at different power settings (Fig. 2A). The variation of the power output, measured over 1 min, was less than 0.13% when measured at 15 mW and a fixed working distance. The percentage variation was lower at a higher power level of 42 mW (SI Appendix, Table 2).

• Download figure
• Open in new tab
• Download powerpoint

Fig. 2.

Characterization of Raman endoscope performance. (A) Power output stability over a working distance of 25 mm away from sample surface. (B) Raman signal reproducibility of the Raman endoscope device over several integration times. Shorter integration times lead to more variability in signal; however, even at 1-ms integration times, the reproducibility is still good with a coefficient of variance (COV) of only ∼2.5%. (C) Raman signal is stable over a working distance of 10 mm away from sample surface. The signal drops off (1/d²) after the Raman endoscope is more than 10 mm away from the sample due to the solid angle collection efficiency. A.U., arbitrary units. (D) Raman endoscope device was able to detect Raman signal above the background threshold level at a depth of 4.5 mm when using 42 mW and a 300-ms integration time and at a depth of 5 mm when using 15 mW with a 1-s integration time in our tissue-mimicking phantom (Right). The phantom was fabricated in-house and was prepared to have similar scattering and absorption properties as human colon tissue. (E) Correlation of the Raman signal to concentration of S440 in a well plate using the Raman endoscope device. The limit of detection was 326 fM (15 mW at 1 s) and 440 fM (42 mW at 300 ms) of SERS nanoparticles in a well plate (Right). The SERS Raman concentration correlates linearly with Raman signal at both laser powers. (F) Linear correlation of the Raman signal to concentration after topically applying diluted concentrations of S440 onto fresh human colon tissue samples (Right). The Raman signal linearly correlates with the concentration of SERS nanoparticles applied.

By performing FVB of the SERS signals dispersed onto the CCD array and using the DCLS algorithm, we were able to obtain reproducible Raman spectra with integration times as short as 1 ms successfully. The reproducibility, measured in terms of the coefficient of variance, further improved for integration times of 10 ms or greater (Fig. 2B and SI Appendix, Signal Stability). Additionally, we confirmed that the Raman signal increases linearly with laser power (SI Appendix, Fig. 3). The majority of the experiments conducted in these studies used either a laser output power of 15 mW with a 1-s CCD acquisition time or a laser output power of 42 mW with a 300-ms acquisition time. These settings were selected to stay within the ANSI MPE limits (SI Appendix, ANSI MPE and Table 1).

The collection efficiency of the multimode fibers in the Raman probe is relatively constant for working distances between 0 and 10 mm (Fig. 2C). In that range, the multimode fibers are being efficiently used for light collection, as verified using ZEMAX (Radiant Zemax, LLC) optical ray-tracing software (SI Appendix, Fig. 4). As the working distance continues to increase beyond 10 mm, the collection efficiency begins to drop as , where d is the working distance between the distal end of the device and the tissue surface with or without Raman nanoparticles (SI Appendix, Working Distance). Even at large working distances, Raman signals were still identifiable (e.g., 40 mm away from the sample surface, which is well beyond the required working distance for applications within the colon). This trend was validated on Raman nanoparticles placed on a quartz slide (SI Appendix, Fig. 5), as well as on human colon tissue biopsies as seen in Fig. 2C.

To evaluate how deep below the surface of tissue the nanoparticles can be imaged, a bench-top phantom was created. The phantom was composed of an agarose solution with similar absorption and scattering properties as the colon tissue (Methods, Depth Study). Using this phantom, a depth of penetration of 4.5 mm was achieved using 42 mW of laser power and 300 ms of acquisition time; using 15 mW and 1 s, respectively, a depth of penetration of 5 mm was achieved (Fig. 2D). More detailed information is provided in Methods (Defining Thresholds and Statistical Analysis).

SERS particles can be engineered to emit different spectral signatures or flavors. For a single SERS particle flavor (e.g., S440) and a fixed working distance, laser power, integration time, and SERS particle volume, it was demonstrated that the mean Raman weighting factor was linearly proportional to the concentration when the nanoparticles were placed in a well plate (Fig. 2E), on a quartz slide (SI Appendix, Fig. 6 A and B), on ex vivo porcine colon tissue (SI Appendix, Fig. 6 C and D), and on human colon biopsies (Fig. 2F). Bench-top experiments using 300 μL of SERS S440 particles placed in a well plate have shown the ability to detect 440 fM SERS particles with a laser illumination of 42 mW, integration time of 300 ms, and working distance of ∼5 mm. This corresponds to a detection limit of 1.3 million nanoparticles, or ∼230 nanoparticles per cell, for an average cell cross-section of 100 μm² (SI Appendix, Detection Limit Calculation). At a laser illumination of 15 mW and a longer integration time of 1 s, we were able to detect as little as a 326-fM concentration of SERS particles. More details are provided in Methods (SERS Concentration Assessment, Mapping and Image Processing, and Well Plate Experiment) and in SI Appendix, Fig. 7.

Multiplexing Experiments.

Multiplexing on a quartz slide.

All multiplexing studies performed herein were evaluated at two different power settings: 15 mW at 1 s per spectral acquisition and 42 mW at 300 ms per spectral acquisition. Due to space constraints, we have reported all the studies performed at 15 mW in SI Appendix. To evaluate our Raman endoscope device’s initial ability to identify and unmix spectra from various SERS nanoparticles correctly, we first placed 10 separate 3-μL samples of each SERS flavor onto a quartz slide. We then mapped the quartz slide containing the 10 individual drops using 1-mm x/y translations at times of either 1 s per spectral acquisition using 15 mW (SI Appendix, Fig. 8) or 300 ms per spectral acquisition using 42 mW (Fig. 3A). The map was analyzed with our postprocessing software, where preassigned reference spectra of each SERS nanoparticle flavor were used to determine the best spectral fit to the measurements in the map. The software sent the intensity maps into 10 separate channels, revealing where in the map each SERS nanoparticle flavor was detected. The resulting image shows all 10 drops correctly separated into their corresponding spectral channel with minimal cross-talk between the channels (Fig. 3A).

• Download figure
• Open in new tab
• Download powerpoint

Fig. 3.

Demonstration of multiplexing on quartz slide at 42 mW. (A) Ten unique flavors of SERS nanoparticles spatially separated onto a piece of quartz. The Raman map acquired identifies all 10 flavors correctly. Notice how each of the flavors is correctly represented in each of the SERS nanoparticle channels in the panels below. (B) Equal mixture of S440 and one other flavor is placed in separate drops across a piece of quartz to characterize dual colocalization of SERS nanoparticles. The Raman map correctly identifies the presence of each of the 2 flavors of SERS nanoparticles in each mixed droplet as shown in the separate SERS channels below. (C) Demonstration of multiple colocalized SERS flavors, including mixtures of 4, 6, 8, and all 10 SERS nanoparticles within the same droplet on quartz. The Raman maps shown depict the correct location of each of the SERS flavors within each of the mixtures. (D) Mixture of 4 SERS nanoparticle flavors, each at varying concentrations. The postprocessing software was able to separate each flavor spectrally into its respective channel correctly, and the Raman maps (Left) show a decrease in Raman intensity that correlates with the concentrations of each of the SERS flavors. (Right) Graph depicts a linear correlation between the SERS concentration of each flavor and the Raman signal, with an R² value of 0.9987.

Next, we assessed the ability of our Raman endoscope device to unmix two colocalized flavors of SERS nanoparticles spectrally. The most abundant flavor we had available was S440; therefore, we mixed equal concentrations of S440 with each of the rest of the nine spectrally unique batches. The resulting intensity maps correctly unmixed each of the nine mixtures containing equal concentrations of S440 and one other SERS flavor (Fig. 3B).

We then demonstrated the ability of our Raman endoscope device to unmix colocalized mixtures of 4, 6, 8, and all 10 SERS nanoparticles successfully at equal concentrations (Fig. 3C). After successfully demonstrating colocalized multiplexed imaging with all 10 SERS nanoparticles, we assessed various mixtures of nanoparticles at varying concentrations. Before deciding which nanoparticle flavors to mix first and at what concentrations, we calculated which SERS spectra were most distinct relative to that of our most abundant flavor, S440, to have the most success with unmixing the spectra. It was determined that S440, S481, S420, and S421 had the most distinct spectra with very little spectral overlap with S440, and each performed very well with one other at varying concentrations. Fig. 3D shows a quantitative linear relationship between each of the SERS nanoparticle concentrations and Raman signal (R² = 0.9987 at 42 mW).

Multiplexing on human colon tissue.

Because our strategy involves a topical application of these SERS nanoparticles onto human tissue for functional tumor-targeting assessment, we evaluated the ability of our Raman endoscope device to multiplex on fresh human colon tissue samples. Each of the 10 nanoparticle flavors was topically applied to fresh human colon tissue samples, and the images revealed a clear distinction between each of the channels with minimal cross-talk (Fig. 4A). Furthermore, equal concentrations of the four most distinct SERS flavors (S440, S481, S420, and S421) were mixed together to assess colocalized multiplexing on human colon tissue, and all four were correctly unmixed in their respective channels (Fig. 4B). Finally, each of these four SERS flavors were mixed together at varying concentrations and topically applied to fresh human colon tissue to reveal a consistent linear correlation between Raman signal and SERS nanoparticle concentration (R² = 0.980 at 42 mW) (Fig. 4C). More information about the 15-mW human multiplex data is provided in SI Appendix, Fig. 9. Therefore, it was determined that the intrinsic background signal from human colon tissue itself did not interfere with the ability of our Raman endoscope device to multiplex effectively in any significant way.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 4.

Demonstration of multiplexing on human tissue at 42 mW. (A) Ten unique flavors of SERS nanoparticles spatially separated onto 10 separate pieces of fresh human colon tissue. The Raman map acquired identifies all 10 flavors correctly. Notice how each of the flavors is correctly represented in each of the SERS nanoparticle channels in the lower panels. (B) Demonstration of colocalized multiplexing, where 4 SERS flavors were equally mixed and applied on a single piece of human colon tissue. The postprocessing software was able to separate each flavor spectrally into its respective channel correctly as shown in the Raman maps (Right). (C) Mixture of 4 SERS nanoparticle flavors, each at varying concentrations (Concen.), was combined and applied to a single piece of human colon tissue. The postprocessing software was able to separate each flavor spectrally into its respective channel correctly. (Left) Raman maps show a decrease in Raman intensity that correlates with the concentrations of each of the SERS flavors. (Right) Graph depicts a linear correlation between the SERS concentration of each flavor and the Raman signal, with an R² value of 0.9796.

Blinded Study: “Scavenger Hunt.”

The purpose of this blinded study was to determine the sensitivity and accuracy of the multiplexed system when put in the hands of a user to detect multiple SERS nanoparticle flavors and mixtures at various concentrations. To conduct this study, a porcine colon was splayed open and a 6 × 10 grid, consisting of 60 boxes, was created on the tissue using suture material (Fig. 5A). Each box was assigned either zero, one, two, three, or four flavors of colocalized SERS particles at a concentration of ∼1 nM. The user was blinded as to what combination of particles was present, if any, in each grid. Within the empty boxes in which no SERS nanoparticles were assigned, a red food dye was placed to ensure that the user was blinded to the presence of the particles. The combinations were randomly generated with a probability that 15–30% of the grid would consist of blanks. The Raman device was mounted in-line with an off-the-shelf semiflexible boroscope (DCS 100; General) to provide the white-light, real-time video guidance (Fig. 5B). The boroscope consisted of a CCD, which transmitted images wirelessly to the computer, and two white-light light-emitting diodes; this was used to help the user navigate over the samples (SI Appendix, Movie S2). Three users were tested in this blind study. A representative graph from each user shows the presence of each SERS flavor within each of the 60 boxes (SI Appendix, Fig. 10). There was a 99% average user accuracy to the presence of the correct combination of SERS nanoparticle flavors, after using a 7× SD threshold (Fig. 5C). The threshold was defined as a multiple of the SD of the background weighting factor, which was added to the mean background weighting factor. No false-negative results were identified throughout the study; therefore, the results were 100% sensitive to the presence of SERS particles across the three users. Additionally, a 96% average specificity was achieved between the three users after applying a 7× SD threshold (Fig. 5C). More details regarding the threshold calculation are provided in Methods (Defining Thresholds). Note that these results reflect the use of SERS nanoparticles at concentrations of ∼1 nM in volumes of 3 μL for each flavor used and that the sensitivity and specificity would presumably change at lower concentrations closer to the limit of detection (subpicomolar). However, to demonstrate the usability and multiplexing capabilities of the Raman endoscope, we chose a dose of ∼1 nM, which has been shown to be nontoxic when administered locally to the colon in mice (18, 22) The sensitivity and specificity were also determined for each SERS flavor combination. More information is provided in Methods (Statistical Analysis) and SI Appendix, Fig. 11.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 5.

Scavenger hunt to demonstrate Raman endoscope usability in conjunction with multiple SERS nanoparticles on tissue. (A) Pig colon tissue configured into a 6 × 10 grid, where users would interrogate each of the 60 squares with the Raman endoscope. Each square was randomly assigned an unknown mixture of SERS nanoparticles, and some squares had no SERS nanoparticles present. (B) Digital photograph of the Raman endoscope setup. Notice how the Raman endoscope is attached to a white light boroscope to help guide the user during the scavenger hunt. (C) Statistical averages across the three blinded users are shown in the table. Because there were no FN results, regardless of the threshold level, 100% sensitivity was achieved. The average user accuracy, also referred to as the CR, is shown as well.

Pilot Clinical Study.

Currently, an IRB protocol has been approved for our Raman device, but the SERS nanoparticles will require further regulatory approval. Therefore, the device alone was evaluated in patients to detect intrinsic Raman signals as potential background for future studies (Fig. 6 A and B and SI Appendix, Movie S3). The purpose of this study was to analyze the variability of the background signals in vivo from the colon wall and the usability of our device.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 6.

Clinical application and utility of the Raman endoscope in patients. (A) Raman endoscope inserted into the instrument channel of a conventional clinical endoscope. (B) Digital photograph taken from the white-light endoscopy component of the clinical endoscope portraying our Raman endoscope protruding from the instrument channel and illuminating a spot on the colon wall in a human patient.

Three male patients undergoing routine screening via colonoscopy consented to participate in this study and were tested with the Raman device by an experienced endoscopist (S.F.). Most endoscopes use a xenon short-arc lamp as the light source, and in our clinical studies, the tower model was an Olympus Evis Exera CLV-180, which uses a 300-W xenon short-arc lamp. The xenon lamp has emission lines in the near IR within our region of interest and contributes to our measured background spectra. The intensity of the lamp, as detected by our endoscope, varies independent of the endoscope’s inherent background signal originating from the laser. By inputting reference signals with the lamp source turned on and off into the DCLS algorithm, we can account for these variations. The robustness of the DCLS algorithm to account for background variation and its ability to cancel out the background signal effectively were assessed in datasets acquired in three patients. For patients 1 and 3, at 15 mW of laser power and 1 s of integration time, the residual signal after background cancellation was 0 ± 35 counts (Fig. 7 A and B). Details are provided in SI Appendix (Clinical Background Cancellation), and results for all three patients are shown in SI Appendix (Fig. 12). When more than one signal from the lamp source was used as a reference in the DCLS algorithm, the background subtraction improved. There are a finite number of additional backgrounds in which a noticeable improvement occurs, after which there is only incremental benefit from adding additional background references to the DCLS algorithm (SI Appendix, Fig. 12E).

• Download figure
• Open in new tab
• Download powerpoint

Fig. 7.

Characterization of Raman endoscope performance in humans. (A) In vivo background signals; the peaks are due to the endoscope light source. Ninety background acquisitions were obtained from patient 1 at 15 mW and 1 s. (Inset) Residual signal after DCLS background cancellation was 0 ± 35 counts. (B) In vivo background signals; the peaks are due to the endoscope light source. The light source unit from patient 3 was different from that of patient 1. Ninety background acquisitions where obtained from patient 3 at 15 mW and 1 s. (Inset) Residual signal after DCLS background cancellation was 0 ± 35 counts. (C) In vivo Raman signals were simulated by superimposing diluted SERS signals acquired on human tissue with the in vivo background signals. The figure shows the superimposed data from patient 3 at 15 mW and 1 s as well as a SERS signal acquired at a concentration of 25 pM on human tissue. (Inset) Extracted SERS signal after DCLS background cancellation. The mean extracted SERS signal, which was also fed through a second-order Savitzky–Golay low-pass smoothing filter to remove high-frequency noise, is shown in black. (D) Superimposed data from patient 3 at 42 mW and 300 ms, and a SERS signal acquired at a concentration of 25 pM. (Inset) Extracted SERS signal after DCLS background cancellation. (E) Figure shows superimposed data from patient 3 at 15 mW and 1 s, and a SERS signal acquired at a concentration of 6 pM. (Inset) Extracted SERS signal after DCLS background cancellation. (F) Simulated data from patient 3 at 42 mW and 300 ms, and a SERS signal acquired at a concentration of 6 pM. (Inset) Extracted SERS signal after DCLS background cancellation. (G) Raman signal extracted from superimposed data at varying concentrations. The signal was evaluated over 90 acquisitions from patient 3 at 15 mW and 1 s. At 6 pM, the signal was well above background. (H) Raman signal extracted from superimposed data at varying concentrations. The signal was evaluated over 90 acquisitions from patient 3 at 42 mW and 300 ms. At 6 pM, the signal was well above background.

Because we were unable to administer our SERS nanoparticles in patients, we simulated Raman signals by superimposing separately measured SERS spectra onto our in vivo patient data. The SERS spectra were obtained by administering diluted S440 nanoparticles onto excised human colon tissue and sampling the spectra with our device. After background subtraction, these spectra were added to our patient acquisitions (in all three patients with the lamp source either on or off) and the DCLS algorithm was used to extract the simulated signal (Fig. 7 C–F). For these simulations, nanoparticle concentrations of 270, 128, 59, 25, and 6 pM were used (SI Appendix, Clinical SERS Simulation and Figs. 13–20 for all three patient datasets with the lamp source either on or off). The Raman signal extracted from the superimposed in vivo data were larger than the background level, and as little as 6 pM was detected in all patients and scenarios (Fig. 7 G and H and SI Appendix, Figs. 13–20). This demonstrates that once the SERS nanoparticles are approved for use in patients, the Raman device would likely be successful at identifying the particles at low concentrations; moreover, our Raman device is capable of doing so even in the presence of additional background noise, such as the xenon lamp.