2.1.

Imaging Setup

The imaging setup is shown in Fig. 1. It is based on a commercial OCT system (OCM1300SS, Thorlabs, New Jersey) and uses a wavelength-swept laser centered at 1325 nm with a bandwidth of 100 nm, a coherence length of 6 mm, and an axial scanning (A-line) rate of 16 kHz. This system was modified to enable HSI by inserting a DCFC in the sample arm. The DCFC (DC1300LEB, Thorlabs, New Jersey) is made by fusing and tapering a DCF and an MM fiber.²⁹ The diameters of the DCF core, inner-cladding, and outer-cladding are 9, 105, and $125\mu \mathrm{m}$, respectively. The MM fiber has a core diameter of $200\mu \mathrm{m}$. The cladding mode extraction efficiency of the DCFC (cladding mode transmission from port S to port B) is higher than 60% while the injection loss of the core signal (core mode losses from port A to port S) is lower than 0.5 dB.

Fig. 1

Optical setup: Ref, reference arm port; Smp, sample arm port; SM, single-mode fiber; MM, multimode fiber; DCF, double-clad fiber; DCFC, double-clad fiber coupler; C1 and C2, collimation lenses; G, dual-axis galvanometers; L, focalization lens; R, retroreflector prism; LF, light guide fiber.

The output of the wavelength-swept laser is sent to an external Michelson type interferometer module (INT-MSI-1300, Thorlabs, New Jersey) that splits the light into the sample and reference arms and also includes a 15-MHz balanced detector. An SMF-28 fiber is connected to the sample arm output of the interferometer module using an FC-APC connector and fusion-spliced at the other end to the DCF input of the DCFC (port A). The DCF output (port S) of the DCFC is then collimated with a 16-mm focal length fiber collimation lens (C2: FC260APC-C, Thorlabs, New Jersey). A dual-axis galvanometer is used to steer the beam (G: GVS002, Thorlabs, New Jersey). An uncoated, 50-mm focal length planoconvex lens (L: LA1131, Thorlabs, New Jersey), placed at one focal length of the galvanometer, is used to focus the beam on the sample for a telecentric scan. The light backscattered by the sample is collected by the fiber core and cladding, but only the signal collected in the core can travel back to the OCT interferometer. Light in the cladding is either transferred to the MM branch of the DCFC or lost at the splice between the DCF and the SMF-28 fiber. The reference arm is composed of an SMF-28 fiber that matches the sample arm dispersion. Since the double-clad and SMF-28 fibers have the same dispersion characteristics, the length of the SMF-28 fiber in the reference arm is simply the sum of all SMF-28 and DCF segments in the sample arm. No dispersion compensation method was used. The interference from the two arms is detected by a dual-balanced detector. The commercial OCT engine performs acquisition, processing, and display.

HSI acquisition is performed using an external white light source (MKII fiber optic light, Nikon Inc., New York). The broad illuminated area is sampled by imaging the face of the DCF (port S) onto the sample using two galvanometer-mounted mirrors. Light reflected from each point is collected primarily by the inner-cladding of the DCF due to the high area ratio between the inner-cladding and the core. Light from the inner-cladding is rerouted to an MM fiber by the DCFC and sent to a spectrometer (Maya2000, Ocean Optics, California). This spectrometer performs continuous 2064-point spectral measurements at a rate of 125 Hz. Despite the spectrometer’s bandwidth spanning from 200 to 1100 nm, the useful bandwidth of the system spans from 450 to 800 nm due to the system’s spectral response and the lamp’s spectrum.

In the current implementation, HSI and OCT acquisitions are performed sequentially and the sample is immobilized to reduce potential motion artifacts throughout the acquisitions. Both OCT and HSI use the same scanning mechanism and share a field of view of $\sim 6\times 6{\mathrm{mm}}^2$. Both the OCT source and the external halogen lamp are active during the acquisition to assess possible cross-talk between modalities. The built-in software of the commercial system is used to acquire the OCT data. HSI acquisition is realized with custom software (LabVIEW, National Instruments, Texas). The OCT volumes are $512\times 512\times 512\text{pixels}$ and the acquisition is performed in less than 20 s. The hyperspectral images are $75\times 75\text{pixels}$ in size over the same field of view and were acquired in 45 s. Additionally, 10 adjacent spectral pixels are binned together to increase signal-to-noise ratio and decrease the total size of the HSI data set, resulting in a spectral resolution of 4.3 nm. The number of pixels acquired in HSI was limited to $75\times 75$ to reduce acquisition time and potential motion artifacts.

2.2.

Image Processing

The acquired reflectance spectra were corrected for background noise and for the spectral response of the system, using Eq. (1)

Beyond the simple appearance of the sample under a certain illumination, the HSI data can be used to perform molecular analysis of the sample. Dawson et al.³⁵ showed, using a simplified skin model, that the logarithm of the inverse reflectance (LIR) is roughly proportional to the sum of the absorbance of the skin constituents [see Eq. (2)]. The same idea was used to highlight molecular constituents in the sample by producing images at wavelengths corresponding to absorption peaks of the target molecule.

2.3.

Combined Optical Coherence Tomography and Hyperspectral Imaging Volumetric Data

The OCT and hyperspectral datasets were combined by overlaying the spectral information in sRGB format onto the surface of OCT slices. To obtain the surface topography from the OCT data, we implemented a custom surface detection algorithm based on the intensity and the first-order derivative of the signal as well as the relative position of the surface pixels to their closest neighbors. Surface detection was performed on the average of three adjacent B-scans further processed using a median filter in the axial direction. For each A-line, the OCT signal above the surface was set to zero. Processing all A-lines in this manner resulted in a topographic map of the sample.

As the HSI and OCT acquisitions shared the same scanning mechanism and field of view and the sample was tightly immobilized throughout the sequential acquisition, the output of both methods were coregistered. After resizing the RGB image produced from HSI data using bilinear interpolation, each individual color pixel was assigned to the corresponding A-line, then 7 pixels just above the surface axial position were colored using the assigned value from the RGB image. The number of pixels overlaid over the surface was chosen to improve visibility while minimizing the loss of topographic information. All algorithmic steps were performed using MATLAB^®.

3.1.

System Performance Analysis

In OCT, the numerical aperture of the beam in the sample space is $\sim 0.04$. This numerical aperture results in a theoretical lateral resolution of $13\mu \mathrm{m}$ (defined as the 10% to 90% distance of the edge response for a Gaussian point spread function). The theoretical axial resolution of the OCT system remains unchanged with respect to the commercial system with a theoretical value of $12\mu \mathrm{m}$ in air. Using the 10% to 90% distance of an edge response, the measured experimental lateral resolution in OCT is $14.5\mu \mathrm{m}$. The experimental axial resolution in OCT is $17.4\mu \mathrm{m}$, measured as the full width at half maximum of an A-line for a perfect reflector. The slight loss in axial resolution could be explained by a small dispersion mismatch between the sample and reference arm induced by the addition of the DCFC.

Assuming a uniform illumination distribution on a flat sample, the theoretical lateral resolution in HSI is given by the diameter of the image of the DCF inner-cladding on the sample and is $\sim 330\mu \mathrm{m}$. The measured experimental lateral resolution for HSI is $345\mu \mathrm{m}$. This resolution was measured using the 10% to 90% distance of an edge response on a sharp interface of a hyperspectral image.

3.2.

Combined Optical Coherence Tomography and Spectral Imaging

A healing epithelial wound on the hand was imaged to evaluate the performance of the system. The subject rested his hand on a microscope slide to help reduce movement artifacts throughout both OCT and HSI acquisitions. The reference arm delay was adjusted to position the objective focal plane just passed the zero-delay plane of OCT. The microscope slide was then placed at the focus by positioning the hand surface a few hundred microns below the zero-delay plane. The two white-light illumination fibers were placed at an angle on each side of the sample to avoid the collection of specular reflection and provide a uniform illumination over the imaged area. A 3-D acquisition was then performed using the OCT engine, rapidly followed by HSI acquisition. OCT and HSI acquisitions could not currently be realized simultaneously due to differences in integration times associated with each modality, a challenge that would be lifted with custom-built systems.

Figure 2(a) shows an image of the healing wound taken with a CCD camera under ambient light conditions. Figure 2(b) displays a typical B-scan of the hand acquired with OCT. A maximum intensity projection was realized to produce en-face images from the OCT C-scan [Fig. 2(c)]. En-face true-color reconstruction was extracted from the HSI data [Fig. 2(d)]. Figure 2(e) plots the LIR representation as a function of wavelength in two regions of interest (ROIs) of the HSI image [shown in Fig. 2(d)] and a typical absorption spectrum of oxygen-saturated hemoglobin.³⁶ Figure 2(e) shows that LIR representation is an adequate approximation to recover the absorption spectrum of the imaged tissue, which is, in this case, mostly composed of hemoglobin. From the absorption spectrum, one could retrieve the relative proportion of the main tissue constituents using spectral unmixing.³⁷ Using the LIR representation, it is possible to detect specific absorption peaks in the sample and assess the presence of a specific absorber. For example, Fig. 2(f) shows the LIR image obtained at 575 nm [vertical line in Fig. 2(e)] which corresponds to an absorption peak in the hemoglobin absorption spectrum and gives insight on the distribution of hemoglobin in the sample. However, one should be careful interpreting Fig. 2(f) as it does not decouple the effect of the sample’s topology on the absorption spectrum and some of the small features observed in Fig. 2(f) could be caused by shadows induced by the irregular surface of the sample.

Fig. 2

OCT and HSI images from a healing epithelial wound: (a) ambient CCD image, (b) typical OCT B-scan, (c) en-face OCT reconstruction, (d) en-face RGB image reconstructed from HSI data, (e) LIR representation as a function of wavelength for two ROIs identified in (d), (f) en-face LIR image at 575 nm, identified as the dashed line in (e). Scale bars are 2 mm. (b) Corresponds to the section-plane represented by the blue line in (c).

Figure 3 shows the combination of OCT and HSI. Figure 3(a) was obtained by overlaying the RGB image [Fig. 2(d)] and the OCT en-face projection [Fig. 2(c)] and adjusting the transparency of both layers. A 7-pixel-wide region just above the surface of an OCT B-scan was colored using the true-color information obtained with HSI [Fig. 3(b)]. The same overlay can be applied in 3-D, as shown in Fig. 3(c).

Fig. 3

Combining OCT and HSI data: (a) OCT en-face projection overlaid with RGB image, (b) OCT B-scan with HSI extracted true-color overlay. An epithelial region is forming in the wounded area (ER); (c) dual-modality 3-D rendering of the hand.