1. Introduction

Planar near-infrared fluorescence imaging (NIRF) is a rapidly growing noninvasive and nonionizing molecular two-dimensional projection imaging technique that has the ability to generate high-resolution, high-sensitivity, and high-contrast images for several preclinical [1] and clinical applications [2]. Even though a large pool of near-infrared (NIR) fluorescence agents is currently under development, indocyanine green (ICG) has been approved by the Food and Drug Administration (FDA) for human intravenous injection for over 50-years in the USA. Owing to its excellent biocompatibility, ICG has been used for a variety of biomedical applications, such as sentinel lymph node mapping in breast cancer [3,4], intraoperative assessment of gastrointestinal anastomotic perfusion [5], tumor margin detection [6], and several applications in plastic surgery [7,8]. Besides, the maximum excitation wavelength of ICG lies in the near-infrared imaging window, in which low tissue absorption and scattering allows a higher probing depth. Furthermore, noise arising from background autofluorescence is minimal within this spectral window. The principle of NIRF with ICG is based on illuminating tissue with an excitation light source, such as light emitting diodes or laser diodes, and detecting emitted fluorescence light with a charge-coupled device (CCD) camera incorporating optical interference filters.

ICG is an amphiphilic cyanine dye whose excitation spectrum depends on the nature of solvent and the concentration [9]. The main peak of the excitation spectrum of ICG shifts from 780 nm to 800 nm when free ICG binds to plasma proteins. Following intravenous injection, the competition between the binding to protein and the aggregation determines the in vivo spectral properties of ICG [10,11]. Some researchers have reported that the slow spectral peak blue shift of 9 nm after intravenous injection is due to specific phospholipid binding [12]. Furthermore, ICG can be utilized as an activatable tumor imaging agent design, which present different spectral properties before and after the activation [11,13–15]. Therefore, spectral investigation, in addition to the fluorescence intensity-based imaging, can help better understand the pharmacokinetics of ICG and its associated biological processes in vivo by providing an additional contrast. Several techniques were previously introduced in order to achieve spectrally-resolved NIRF such as using liquid crystal tunable filters [16] or a spectrograph [17]. However, these techniques were instrumentally implemented at the detection side and limited to either improve the spectral-resolution or the spatio-spectral scanning speed. In this paper, we propose a novel method that provides a new additional contrast that can be leveraged to improve the standard two-dimensional fluorescence intensity imaging. The method is based on the acquisition of spatio-spectrally resolved NIRF images by employing a near-infrared wavelength-swept laser and a principal component analysis (PCA)-based image processing technique.

2. Material and methods

Several research groups have previously reported use of near-infrared wavelength-swept lasers for biomedical imaging [18,19]. Meanwhile, we developed a wavelength-swept laser centered at 800 nm to image micro-structural and functional information of blood in vitro [20], high-resolution optical coherence tomography depth-resolved retina layers [21], and real-time surface plasmon resonance imaging [22]. Our new wavelength-swept laser is able to tune its wavelength from 784 nm to 820 nm, which covers the maximum excitation wavelength of ICG. It is equipped with a holographic transmission grating and a Galvo-scanner used as a wavelength-selector in order to avoid the thermal instability caused by piezoelectric transducer-based wavelength-selector standardly used, Fig. 1(a). A traveling-wave semiconductor optical amplifier (SOA) is employed to provide the optical gain from the spectral range of 784 nm to 820 nm. A 10-dB fiber-optic directional coupler is connected with the output port of the SOA. A 90% arm of the optical directional coupler is used to drain the laser output while the other 10% arm is coupled to the external fiber-optic laser cavity to feedback stimulated light towards the traveling-wave SOA. The wavelength-selector consists of a holographic transmission grating, a reflection mirror, and a Galvo-scanner mirror.

 

Fig. 1. (a) Schematic of the wavelength-swept laser and our wide-field NIRF system. SOA: semiconductor optical amplifier, P: polarization controller, CR: circulator, CL: collimator, GS: Galvo-scanner, G: grating, M: mirror, C: coupler, IS: isolator, Sp: splitter, F1: exciter bandpass filter, D: diffuser, F2: emitter bandpass filter, CCD: charge coupled camera, L: lens. (b) Normalized output spectra of the wavelength-swept laser at representative tuning wavelengths. The passband of the filters is highlighted for the (orange) exciter F1 and (green) fluorescence collection F2. (c) Measured wavelength-swept laser output power. The representative wavelengths presented in (b) are shown with blue triangle markers.

Download Full Size | PDF

The wavelength selection is made by rotation of the Galvo-scanner mirror such that a narrow-bandwidth of the light within the numerical aperture of a fiber-optic collimator can act as feedback to the external laser cavity. The measured maximum output power and the steady-state -3 dB laser linewidth are 14 mW and 0.2 nm, respectively at 800 nm. The normalized steady-state spectra of our laser at five representative wavelengths are shown in Fig. 1 (b).

One can notice the important overlap of the laser spectral tails at the longer wavelengths, which degrades the performance of the fluorescence rejection filters. Therefore, a bandpass filter F1 (FF01-794/32, Semrock Inc., USA) is used at the output port of the wavelength-swept laser to reduce these spectral tails and thus reduce excitation light leakage. To maximize the out-band rejection performance of the filter F1, light is first collimated prior the filter and then focused into the fiber again using compact fiber optic collimators. The output of the laser is connected to a 1-to-2 fiber optic splitter (90/10). The 10% arm of the fiber optic splitter is connected to a commercial spectrometer (USB 2000+, Oceanoptics Inc., USA) for laser spectrum, Fig. 1(b), and power measurements, Fig. 1(c). The 90% arm output is homogenized by an optical diffuser (EDC-20-E, RPC Photonics Inc., USA) and used for wide-field illumination of the surface of the phantom with a power density of approximately 0.5 mW/cm². A schematic of the system implementing epi-illumination scheme is shown in figure1(a). The diameter of illumination spot on the surface is 60 mm. A cooled CCD camera (ColdBlue, PerkinElmer optoelectronics, USA) with an imaging lens (Marco F/2.8, Sigma Inc., USA) is used to collect fluorescence images at the surface of the phantom [11]. We employed two-cascaded bandpass fluorescence interference filters F2 (830-10, MKphotonics Inc., USA) between the lens and the camera to selectively collect the fluorescence emission light and reject the excitation laser light, Fig. 1(a-b).

The performance of our spectrally-resolved NIRF approach is evaluated using a 60 mm x 46 mm x 30 mm rectangular cuboid agarose phantom, Fig. 1(a). The optical properties of the phantom are adjusted using Indian ink and Intralipid to mimic the absorption and reduced scattering coefficients of biotissue and are set to 0.01 mm^-1 and 0.8 mm^-1, respectively. Two glass tubes (diameter: 3 mm, wall thickness: 0.3 mm, length: 80 mm) are imbedded 5 mm below the surface of the phantom and positioned 10 mm apart. One tube is filled with ICG dissolved in distilled water with a concentration of 3.78 µM. The second one contained ICG dissolved in dimethyl sulfoxide (DMSO) with a concentration of 0.72 µM. The difference in concentration is intentionally chosen to compensate for the higher quantum yield of fluorescence of ICG dissolved in DMSO.

3. Results

The spectra of both solutions are measured using the commercial spectrometer, Fig. 2(a). The absorption maxima of ICG in water and in DMSO are found at 783 nm and 795 nm, respectively. Henceforward, the solution containing ICG dissolved in distilled water and ICG dissolved in dimethyl sulfoxide (DMSO) will be respectively referred to as ICG and DMSO. The NIRF images were collected using a set of twenty wavelengths (N = 20) of the wavelength-swept laser from 784 nm to 803 nm with a 1 nm increment. In addition to using filter F1, the highest wavelength is set to 803 nm since higher wavelengths show a strong spectral component that overlaps with the bandpass of F2 resulting in a significant excitation light leakage. In this spectral band, the chosen wavelengths have lesser and relatively similar tails. Consequently, their leakage contribution to the measured fluorescence images is nearly equal, and thus do not affect the performance of the data processing technique.

 

Fig. 2. (a) Normalized absorption spectra of ICG dissolved in (blue) DMSO and (red) water. The red shaded region depicts the used wavelengths range. (b-left) Ambient-light top view of the phantom. The boundaries of the phantom are delineated with a red dash-dot line. Processing ROIs are delineated with dash line circles (blue) DMSO and (red) ICG. (b-right) The corrected fluorescence images acquired using the laser wavelengths, presented in Fig. 1(b). (c) Mean fluorescence intensity within the ROIs and their 2nd order polynomial fits.

Download Full Size | PDF

Our CCD camera allows to acquire 2280 pixels x 1528 pixels images. The pixels are binned during acquisition to increase the signal strength using a factor of 4, resulting in 570 pixels x 382 pixels images with a pixel size of 0.25 mm × 0.25 mm. Each image is acquired using an integration time of two seconds. To account for the laser output power variability at different wavelengths, each fluorescence image is normalized by its corresponding laser output power, Fig. 1(c). Corrected fluorescence images at five representative wavelengths are shown in Fig. 2(b). Technically, fluorescence intensity is proportional to the quantum yield of fluorescence, the concentration of the fluorophore, and the absorbed energy during excitation. Considering that the quantum yield and the concentration of the fluorophore are constant during the imaging session, the variation in the fluorescence intensity only results from the difference in the excitation wavelength. Mean fluorescence intensity is calculated over two regions of interest (ROIs) defined above the two tubes, Fig. 2(b). Second order polynomial fits performed on the obtained multi-wavelength mean fluorescence intensities show that fluorescence emission of the two solutions does not behave similarly across the used spectral window, Fig. 2(c). Since excitation light leakage is similar at all wavelengths, the difference in the wavelength-dependent variation of the measured fluorescence signal is solely due to the difference in the absorption spectrum of the two solutions, Fig. 2(a). This observable difference demonstrates the ability to spectrally distinguish between photons emitted from ICG dissolved in two different solvents. However, obtaining spatial information requires application of this approach pixel by pixel, which is very time consuming. A more elegant way to assess the entire spatio-spectral fluorescence data is to extract pixel-specific and wavelength-dependent signatures by using principal component analysis (PCA) [23–30]. PCA is performed using the “pca” MATLAB function, on the multi-wavelength fluorescence images within a given analysis region. Here, this region is defined as the set of pixels having a fluorescence signal intensity higher than 10% of the maximum fluorescence intensity in all images. The principal components (PCs) are calculated and used to project the multi-wavelength fluorescence data into the principal component (PC) space and generate a PC scatter plot of the PC feature values, which will be referred to as PC henceforth, Fig. 3(a). Considering the high-sensitivity of PC3 to noise and its disability to distinguish between the wavelength-dependent light profiles emitted from the two different solutions, only the first two principal components (PC1 and PC2) are calculated and used to project the multi-wavelength fluorescence data into the PC space [24,26] . Each point in the PC scatter plot represents the wavelength-dependent fluorescence variation at a specific pixel. In addition to the position within the PC scatter plot, the PC values are used to generate a red-green-blue (RGB) display color for each of these points. The color code for each point is defined by attributing its PC1 and PC2 values to the blue and red channels and setting the green channel to zero, RGB = [PC2, 0, PC1], Fig. 3(a). Note that the PC values were first converted to gray scale using the “mat2gray” MATLAB function then normalized to 255. As can be seen on Fig. 3(a), two edges of the point cloud can be distinguish based on their PC1 and PC2 values, as well as based on their color.

 

Fig. 3. (a) Scatter plot of PC1 and PC2 feature values obtained on the pixels within the processing region. The color of each point is set as RGB = [PC2, 0, PC1]. (b) The pseudo color-coded image PC-RGB obtained by attributing each point color in (a) to its position in the image space superimposed on the ambient-light image of the phantom. (c) Scatter plot of PC1 and PC2 feature values obtained on the pixels within the processing region. The color of each point is set as blue for DMSO and red for ICG. The brightness of the points in each cluster is proportional to their distance from the center of their cluster. (d) The pseudo color-coded image k-means-RGB obtained by attributing each point color in (c) to its position in the image space superimposed on the ambient-light image of the phantom.

Download Full Size | PDF

To directly visualize this result in image space, the color obtained at each point in the PC scatter plot is assigned to its corresponding pixel position in the image space [24]. The resulting image (PC-RGB) is superimposed on the gray scale ambient-light image of the phantom, Fig. 3(b). The spatio-spectral information provided by the PC-RGB allows to directly visualize and distinguish between the two tubes. Considering our pca-based data processing time of ∼110 ms on a Dell Precision 5820 Workstation (Intel Core Xeon Processor W-2255 CPU 4.7 GHz, ten cores, 64 Gb memory), this method can be used as a fast tool when near-real-time imaging is required such as during intraoperative procedures. However, although the colors right above the tubes are very distinct, they remained relatively similar elsewhere. Thus, additional processing is required to obtain a total distinction between the photons emitted from one tube or the other. In the image space, performing this processing is not straightforward in the region between the two tubes. Therefore, we propose to perform this step in the PC space. Indeed, a total distinction between the photons emitted from one tube or the other corresponds to separating the PC scatter plot into two point-clouds. Here, we utilize the k-means clustering algorithm [31] to separate between the two point-clouds, Fig. 3(c). First, blue and red colors are respectively attributed to the DMSO and ICG point-clouds for consistency with the PC-RGB method results. Second, the brightness of each point was proportional to its distance from the center of the cluster to which it belongs. Thus, two information is obtained through this representation: 1) the color of the point defines the type of the solution in the tube, 2) the brightness of the point shows the distance of the corresponding pixel from that given tube.

Similar to the PC-RGB method, the resulting color at each point in the PC scatter plot is assigned to its corresponding pixel position in the image space [24]. The resulting image (k-means-RGB) is overlayed on the gray scale ambient-light image of the phantom, Fig. 3(d). The spatio-spectral information provided by the k-means-RGB allows to accurately distinguish between the two tubes. Considering the additional k-means algorithm running time of ∼3.7 seconds, this method will not allow a fast visualization but can be considered complimentary to the PC-RGB method. In fact, it can be used when total distinction between the two regions is needed such as during tumor resection. Nevertheless, the emergence of novel parallelized k-means algorithms as well as the availability of powerful graphic processing units (GPU) will enable near real-time k-means-RGB utilization [32,33].

4. Discussion and conclusion

We demonstrated the proof-of-concept of a new spatio-spectrally-resolved NIRF imaging technique able to resolve ICG in different media using a wavelength-swept laser. Utilizing a swept-wavelength laser, this method provides a new fast and sensitive way to extract spectral information based on the absorption spectra of the fluorophores in addition to the conventional intensity-based fluorescence imaging. This additional information obtained from the spectral dimension can be leveraged in various applications, such as tumor heterogeneity imaging. Indeed, ICG present in different regions of a tumor might exhibit spectral variations based on tumor microenvironment heterogeneity. Thus, in contrast to conventional intensity-based imaging, the spectral information obtained with our technique might be a powerful tool to image and reveal tumor heterogeneity. Other exciting applications might be investigating tumor microenvironmental changes in vivo immediately following treatments such as radiation therapy [24]. Also, our method is dedicated to performing in the near infrared spectral range, where the tissue autofluorescence is neglectable. Thus, the acquired fluorescence emission signals are not degraded by any other component besides the ICG emission. However, the presence of multiple near infrared agents in the field of view will require developing more sophisticated algorithms such the ones using demultiplexing libraries to separate their individual contribution [34].

However, our new technique still poses practical limitations. First, the power of this wavelength-swept laser version is relatively low. Second, a compromise has to be made between the imaging speed and fluorescence signal levels due to the low efficiency of the camera. To overcome these limitations, our next generation system design will utilize an optical amplifier and a high-gain and high-speed CCD camera, such as an intensified CCD (ICCD) or an Electron-Multiplying CCD. Preliminary bench tests using a demo ICCD showed that images with similar SNR can be obtained with an integration time of only ∼30 ms. Moreover, the number of wavelengths at which measurements have been performed (N = 20) has been heuristically chosen. Similar results can be obtained using a smaller number of measurements at optimally chosen wavelengths. Therefore, optimizing the number of wavelengths as well as implementing the new generation instrumentation will allow fast spectrally resolved fluorescence imaging. Indeed, considering our pca-based data processing time of ∼110 ms, using approximately ten wavelengths with ICCD integration time of ∼30 ms, would result in capability of imaging at two multispectral images per second to the end user.

This technique is not limited to differentiation of ICG states but can also be a promising method for simultaneously imaging multiple near-infrared fluorophores and activatable fluorescence agents for specific biomolecule targeting.