1. Introduction

Light scattering techniques continue to be developed for diagnosing disease [1]. Recent applications have focused on diverse applications such as identifying malignant pancreatic cysts [2], non-invasive diagnosis of bladder cancer [3], screening for leukemia in peripheral blood [4], and detection of lung cancer via nanoscale chromatin alternations [5]. Angle-resolved low-coherence interferometry (a/LCI) is a light scattering technique that yields depth-resolved maps of nuclear morphology within epithelial tissues [6]. By capturing backscattered light across various scattering angles, a/LCI enables the determination of nuclear size and density - crucial biomarkers for identifying dysplasia. It produces high sensitivity in identifying early neoplastic changes in diverse anatomical sites, such as the cervix [7,8], esophagus [9–11], and colon [12].

However, the acquisition of angle-resolved light scattering signals from cell nuclei, the core of a/LCI, poses challenges due to instrument complexity. Conventional a/LCI setups involve multi-element fiber bundles and polarization-maintaining (PM) delivery fibers linked to a distal optic, often a gradient-index (GRIN) lens. The GRIN lens collimates the obliquely angled illumination beam onto the sample, mapping angle-resolved scattering to specific fiber bundle elements. This format requires several constraints, including: achieving precision optical pathlength matching between fibers of the bundle (<500 µm over a 2-meter probe), providing sufficient flexibility for endoscopic use, delivering in a package suitable for use during endoscopy, all while permitting high optical throughput and allowing detection of the fundamental optical mode while suppressing additional higher order modes often present in fiber bundles. An integrated solution that can address these challenges remains elusive [13].

Recent efforts have suggested new approaches for developing robust, low cost fiber optic probes for coherence imaging [14–16]. Historically, a/LCI fiber optic probes employed leached imaging fiber bundles (Schott, NA) to transmit angular scattering data from tissue surfaces to the optical engine [17]. However, these bundles are not widely commercially available at lengths required for clinical endoscopy applications. When they are, they are expensive (> $\$ 1,000/{\textrm{meter}}$), exhibit considerable transmission losses (10-20% per meter in near-infrared), and fragility due to the removal of the cladding material between individual borosilicate elements, each measuring only 6-12 µm in diameter [18]. Their performance for interferometric applications is also not ideal, as they are not intended to operate as single-mode waveguides. More affordable alternatives demonstrate poor pathlength stability [13] and restricted flexibility, rendering them unsuitable for coherent imaging. A novel solution is essential.

In response, we introduce a pathlength-matched linear fiber array (PLFA) crafted from economical single-mode fibers and a 3D-printed ferrule. The PLFA enables single-mode propagation, ensuring high throughput (> 99.93%/meter), flexibility, and cost reduction compared to coherent fiber bundles. The distal optics of the a/LCI probe are encompassed within a custom 3D-printed enclosure designed to be compatible with the dimensions of a standard GI endoscope. Notably, this marks the first a/LCI device to enable side viewing, offering new possibilities for clinical application. In addition to a/LCI, this device provides simultaneous endoscopic optical coherence tomography (OCT) for esophageal mapping and metaplastic Barrett's esophagus (BE) identification. These two modalities are complementary - OCT provides real-time feedback on tissue architecture, while a/LCI quantifies nuclear morphology, providing an accurate biomarker for dysplasia detection. The side-viewing a/LCI probe was integrated with a rotational endoscopic OCT system [19] designed for routine esophageal epithelium imaging during upper GI endoscopy. This combined modality, encapsulated in a slender paddle similar in form factor to a radiofrequency ablation paddle [20], is fabricated using a 3D-printed biocompatible resin, providing structural robustness, water resistance, and tissue apposition. Its functionality in both a/LCI and OCT imaging is verified here using both imaging phantoms and human subjects. Hardware and software components are integrated into a mobile cart for deployment in the clinic. Custom software coordinates both a/LCI and OCT systems while furnishing clinicians with an intuitive interface for use during endoscopic procedures.

2. Methods

2.1 Pathlength-matched linear fiber array

The design of the pathlength-matched linear fiber array (PLFA) is shown in Fig. 1. The PLFA consists of 30 single-mode fibers aligned in a linear configuration and matched in optical pathlength to very high precision (<50 µm). A spectral-domain Mach-Zehnder interferometer, previously used to characterize fiber bundles [13], was used to precisely characterize the optical pathlength of each SM fiber to a sensitivity of ∼5 µm. Thirty single-mode fibers (SM800G80, Thorlabs) with 80 µm cladding to improve lateral packing, were rough-cut to ∼2.5 meters manually. An identical fiber, approximately 1 mm shorter in length, was connectorized and jacketed for use as a pathlength standard. A compensating fiber was placed in the interferometer to achieve interference on the detector.

 

Fig. 1. Construction of the PLFA. (A) Ferrule and associated fibers before application of UV epoxy. Scale bar = 10 mm(b) Closeup of the ferrule showing insertion of the fibers. (c) En face image of the ferrule, showing fiber packing. Scale bar = 5 mm The two distal elements are polarization-maintaining (PM) fibers, oriented so that the slow axis is aligned with with the fiber axis. Scale bar = 0.5 mm

Download Full Size | PDF

Once the precise optical pathlength of the reference fiber was noted, each bare fiber was loaded into the interferometer for pathlength inspection and matching. The interferometer is a spectral domain scheme where pathlength mismatch between the two arms appears as an oscillation across the spectrum. The spectrum is then processed similar to OCT schemes to produce a depth resolved profile. Bare fibers under test were connectorized using two 80 µm bare fiber adaptors (Precision Fiber Products, Inc.), such that the projected end of the fiber past the adaptor tip was just visible to the naked eye. From here, the fiber was briefly air-polished to remove the projecting portion and polished on a tower (KrellTech) for ∼15 seconds using a 6 µm grit polishing film, creating a surface smooth enough to allow light to propagate through. The bare fiber was then inserted into the interferometer to examine its optical pathlength. The difference in pathlength between the fiber under test and the reference fiber can be determined by observing the peak in the depth resolved profile. If the test fiber was found to be too long, it was removed from the interferometer, and again a small portion just visible to the naked eye was pushed out of the adapter and re-polished. It was found empirically that this process can remove optical pathlength in increments of 30-70 µm. Polishing and pathlength inspection were repeated for 30 fibers, until each 2.5-meter fiber was matched to within ±50 µm, representing a normalized length precision of ±Δl/L = 0.002%. Each fiber was marked near the tip with black ink in case a breakage rendered it unusable and set aside for later use. The process for pathlength matching each fiber was approximately 15-20 minutes for a trained operator.

To assemble the fibers into a linear array, a ferrule was designed for tight lateral packing. On the distal end, the ferrule includes thirty path matched single mode fibers, as well as two illumination fibers. On the proximal end, the ferrule only included the linear array because the illumination fibers were separately connectorized for interfacing with the a/LCI system. To avoid expensive precision-machined parts, the ferrule was designed and printed on a stereolithography-type 3D printer (Form2, Formlabs) which exhibits excellent z-resolution of 25 µm. Two ferrule halves (4.4 × 2 × 10 mm) were designed and printeded with a 50 µm by 3.2 mm depression on one side, which corresponded to a 100 µm by 3.2 mm linear slot once the two halves were aligned. Two access channels (Fig. 1(A)) were added to one side of the ferrule to help pack down the fibers during assembly. The ferrule halves were held together with a thin strip of adhesive tape (and placed in a fixture for fiber insertion (Fig. 1(A), B). A glass coverslip was placed about 1 mm from the front of the ferrule to act as a backstop for the fiber, and an inspection camera was positioned on the other side of the coverslip to monitor the fiber placement process. Figure 1 C shows the end face of the ferrule with fibers inserted.

In a/LCI, it is important to control the polarization of the illumination beam [21]]. To that end, the first fiber inserted into the ferrule, which provides the illumination beam, was a polarization-maintaining (PM) fiber (HB800 G, Thorlabs), which was polished using a bare-fiber adaptor to reveal the stress rods. The PM fiber, approx. 30 cm longer than the single-mode fibers, was inserted into the ferrule, and rotated so that the slow axis was aligned with the fiber array. We found empirically that the rotation of the PM fiber could be controlled manually with an accuracy of <5° [17].

The distal end of the array was assembled first. One PM fiber was inserted, followed by each of the 30 pathlength-matched fibers and finally an additional PM fiber was inserted and aligned at the opposite end. This second PM fiber could be used for illumination in case the primary PM fiber failed mechanically during assembly. Once all fibers were inserted, a single drop of low-viscosity optical adhesive (Norland NOA 72) was placed into the access channel of the ferrule and allowed to spread between the fibers. The adhesive was spot-cured using a UV light source (LOCTITE EQ CL32), and additional epoxy was added and cured to protect the stripped ends of the fibers proximal to the ferrule. The proximal end of the PLFA was then assembled by inserting the other ends of the 30 fibers in the same order (excluding the two separately connectorized PM fibers).

Once the fibers were assembled, polyolefin heat shrink tubing was used to create a kink-proof protective jacket around the loose fiber between ferrules. Short lengths of ¼” tubing (3:1 shrink ratio) were applied at the fiber insertion point for strain relief. The entire assembly was then threaded through a ¼” heat shrink tube using a pull-through method, heated, and reinserted into a 3/8” tubing, which was also heated. The resulting jacket was ∼4.7 mm in diameter, and flexible, but rigid enough to prevent kinking. The surface of each ferrule was polished using a custom clamp attached to the polishing tower, with extended polishing times used to create an even grind. We polished the ferrules using a glass pad instead of the rubber backing of the standard connector polishing pads to ensure a flat ferrule surface. The finished PLFA was characterized for optical pathlength precision and fiber placement.

2.2 a/LCI Distal Optics

In contrast to earlier a/LCI setups that employed a quarter-pitch GRIN lens in direct contact with the fiber bundle [17], our Pathlength-Matched Linear Fiber Array (PLFA) had a width which slightly exceeded 2 mm, surpassing the limit of commercially-available GRIN lenses with a maximum diameter of 2 mm. This factor, combined with the desire for a side-viewing configuration optimized for endoscopy, necessitated an alternative approach. We devised a miniature 4f imaging system designed to relay and demagnify the fiber plane onto the back focal surface of a GRIN lens (Fig. 2(A)) while also enabling a 90° turn in the optical train. The 4f configuration comprised an achromatic lens (L2, Edmund Optics, 5 mm diameter, f15 mm) and a molded aspheric lens (L1, Thorlabs N414-B, f3.3 mm), resulting in a magnification factor of 0.22. Notably, prior a/LCI devices were forward-viewing [8,17], whereas a side-viewing configuration was desired, as it was deemed to offer more utility for upper endoscopy. To achieve this orientation, a gold-coated right-angle prism (P,Thorlabs, 3 mm) was introduced following the 4f system, redirecting the optical axis by 90° toward a GRIN lens (Edmund Optics, 1.0 mm diameter, 0 mm working distance). The 4f system created an image of the fiber array at the back surface of the GRIN lens (G), while the GRIN lens itself collimated the illumination beam onto the sample at an oblique angle of approximately 15° to the optical axis. A round glass coverslip (#0 thickness) is placed between the GRIN lens and the sample to protect the optic and probe. Scattered photons from the sample were collected by this GRIN lens and relayed back to the PLFA, such that each fiber would collect a distinct scattering angle.

 

Fig. 2. Assembly of the multichannel a/LCI-OCT probe body in a dual-modality paddle housing both a/LCI and rotational OCT components. (A) Lens types and spacings were optimized in Zemax OpticStudio (Zemax, LLC), including prism P, GRIN Lens G, molded aspheric lens L1 and achromatic lens L2, and (B) a housing with slots for the optics was designed in SolidWorks (Dassault Systèmes), with sawtooth mating features for improved bonding and fixed cuff attached to the endoscope. Scale bar = 5 mm (C,D) The internal features are designed to perfectly fit the micro-optics, with no epoxy required. Location of G, L and PLFA are indicated. Scale bar = 5 mm (E) The GRIN lens is the smallest and most sensitive element to misalignment. Tiny ledges are printed to seat the GRIN lens in the correct focal plane. Scale bar = 1 mm(F) The assembled and polished probe is only 4.5 mm in thickness, ramping up to 5.3 mm near the imaging tip and (G) similar in form to a radiofrequency ablation paddle. Red arrow indicates a/LCI probe, blue arrow indicates location of OCT window. Scale bar = 5 mm.

Download Full Size | PDF

Analysis of this optical train in Zemax OpticStudio (Zemax, LLC) shows that the imaging characteristics align with the design goals. The primary source of aberration in the system is the asphere (L1), however, the most significant aberrations (spherical, coma) induced by this optic exhibit opposite signs on each of the two lens surfaces, and largely cancel. A small amount of distortion is introduced by the interface between the GRIN lens, G, and coverslip, but this is a minor effect. Importantly, the collimation error of the illumination beam, which could theoretically limit the angular resolution of the system, is approximately ∼0.2 °. This is much less than the angular sampling of the PLFA in our design, ensuring that aberrations do not limit our sampling of the angular scattering function.

To further reduce form factor, lens truncation (the term for removing a chord from the circular lens aperture) was employed by cutting 1 mm of material from the top and bottom of each lens in the 4f system using a custom lens grinding fixture. This was possible due to the nature of a/LCI, which collects scattered light along a single dimension of the optical aperture, reducing aperture requirements in the other dimension. Lens truncation served to reduce the thickness of the paddle by 2 mm, a substantial amount considering the small form factors required for devices attached externally to an endoscope.

2.3 Dual-modality imaging probe design and construction

To accommodate both the miniature angle-resolved low-coherence interferometry (a/LCI) probe with truncated lenses and the rotational optical coherence tomography (OCT) device, we designed a dual modality paddle. The paddle's foundation features internal spaces tailored to house the micro-optics described above including the truncated lenses, effectively minimizing the paddle's thickness. Figure 2 (B,C) showcases the finalized probe, both in Computer-Aided Design (CAD) and as a physical print (Fig. 2 D) with integrated lenses. Figure 2 E shows a closeup detail of the prism and imaging optics indicated by the dashed box in Fig. 2 D. Figure 2 F shows an image of a printed and assembled probe head with both a/LCI and OCT channels indicated. The PLFA ferrule described above is contained in the white tubing entering the left end of the paddle.

The two imaging channels were arranged side by side to ensure proximity between the a/LCI and OCT ports (detail shown in Fig. 2 G) and establish measurement parity between the modalities. Insights from our clinical collaborators indicated that dysplasia tends to manifest within tissue on a scale of approximately 5 mm at minimum. This insight guided our design to ensure imaging fields were sufficiently close. In the completed device, the center of the a/LCI field of view lay approximately 4 mm from the center of the OCT field of view and within 3 mm of the nearest imaging point. This arrangement ensured that tissue features captured by one modality closely aligned with those captured by the other.

The paddle thickness was a pivotal parameter. In its final form, the paddle measured 4.5 mm in thickness across most of its body, with a slight extension to 5.3 mm near the imaging point. This thickness profile was akin to several commercial devices employed for focal radiofrequency ablation [22], and notably slimmer than a previous proof-of-concept probe which measured 9.6 mm thick prior to an optical redesign and lens truncation. A complete description of the design history of the probe, including optical, mechanical, and material considerations, may be found online [23].

Specific considerations were taken into account when designing the imaging windows. While a/LCI can utilize a small, uncoated glass coverslip, OCT windows generally require an anti-reflection (AR) coating to mitigate common path interference in the detected signal. To meet this requirement, a sapphire window (Precision Micro-Optics Inc.) sized at 6.5 × 5.0 × 0.3 mm was designed, incorporating an AR coating that provided less than 0.25% reflectance at 1310 ± 20 nm. An indentation for this window was included in the paddle's design, simplifying its integration and subsequent sealing with biocompatible optical adhesive (Norland). The comprehensive design and assembly details of the dual-modality paddle are visually presented in Fig. 2.

To hold the optics in place, a probe body (Fig. 2(C)) was designed to tightly fit the contours of each lens and prism, such that each optical element would fit without the need for epoxy. This was achieved in practice by adding a 50 µm buffer to the exterior of each lens CAD file, and constructing a housing around the resulting surfaces. The probe body was designed in two parts for easy assembly, with sawtooth mating features to increase the surface area between halves for stronger bonding and easy alignment. A lip was added to support the GRIN lens on a portion of the proximal surface not used for imaging, and extruded cuts were added near the corners of the prism to account for rounding of sharp features in the 3D printing process. To hold the PLFA, a slot was added to the proximal end of the probe, with four 2.26 mm holes that could be threaded with a 4-40 tap to hold the fiber array in place using set screws. The probe was printed on a stereolithography 3D printer (Form 2, Formlabs with later designs using Form 3) using a biocompatible dental resin. Optical elements were carefully slotted into place, with only a small bit of epoxy used to hold the prism in place. Once assembled, the triangular mating features were painted using the same resin used to print the paddle, and held together in a vice for 24 hours until set. Finally, a 3 mm circular glass coverslip (No. 1 thickness, Thomas Scientific) was bonded to the front surface of the GRIN using an optical adhesive (Norland, NOA81) to provide protection for the surface. The PLFA was then inserted into the slot and adjusted until the output beam was collimated, and then fixed in place using set screws. Once set, the screws were coated in a biocompatible, waterproof UV-curing resin, and then overcoated with the biocompatible dental resin to create a waterproof seal around the device. Figure 3 presents a graphical overview of the probe fabrication process.

 

Fig. 3. Flow chart of PLFA Assembly.

Download Full Size | PDF

2.4 Integrated dual modality optical engine

The dual modality imaging system combines an OCT system with the a/LCI engine (Fig. 4). A full description of the OCT system used was first reported by our group in Chu et al [19]. To summarize, the OCT system utilized a spectral-domain fiber-optic Michelson interferometer design with a superluminescent diode (SLD) centered at 1318 nm and a 3 dB bandwidth of 83 nm for illumination. The SLD provided an output of 20 mW but only a fraction of this power reached the tissue. The sample arm employs a fiber-optic rotary junction (FORJ) to enable distal optics rotation. The FORJ was mechanically driven by a stepper motor, and both mechanically and optically coupled to the probe via a rectangular SC-APC connector. The probe shaft was comprised of a three-layer wound steel torque coil encasing an SMF-28 fiber. The probe distal optics were produced by a commercial vendor (GRINtech, Jena Germany) and consisted of a GRIN objective lens and a right-angle prism, creating a focused beam at 90 degrees to the optical fiber axis for side-viewing during tissue imaging. The probe rotated at 13.9 revolutions per second. Returning light was combined with reference light and detected using a custom spectrometer based on previous designs [24]. The data are processed in real time to present continuous OCT images to the instrument operator.

 

Fig. 4. Schematic diagram depicting the combined OCT and a/LCI optical engine with associated probe. OCT: The OCT system employed a spectral-domain fiber-optic Michelson interferometer design with a superluminescent (SLD) diode centered at 1318 nm, utilizing a fiber-optic rotary junction to rotate the distal optics and enable side-viewing. a/LCI: Light from the SLD is split into sample and reference arms in a ratio of 99:1. Sample light was passed through a polarization controller and coupled to the PM fiber in the pathlength matched linear fiber array (PLFA). Scattered light is collected by the path-matched SM fibers and delivered to the slit of a custom imaging spectrometer, along with a collimated reference beam. S1, S2 represent optical shutters, DAQ is a data acquisition module, PC is a polarization controller.

Download Full Size | PDF

The integrated a/LCI optical engine was designed to detect scattered photons collected by the probe. Light from a superluminescent diode (SLD, ${\lambda _0}$ = 830 nm, 20 mW) was first split using a 99:1 fiber splitter (AC Photonics, Inc) into sample and reference arms. The sample arm light passes through a manual polarization controller (Thorlabs) and is coupled into one of the PM fibers of the PLFA. The illumination beam from the PM fiber is collimated by the probe optics, but since it was offset from the optical axis, it illuminates the sample at an oblique angle. Scattered light is collected and transmitted back down the PLFA, with each fiber collecting light scattered at a distinct angle. The sample light is combined with a reference field at a beam splitter and the combined fields are imaged onto the slit of an imaging spectrometer using a 4f lens system. A ∼5 meter reference fiber is included in the reference arm to achieve path-length matching with the fiber probe and produce interference on the detector.

The imaging spectrometer was designed specifically for this probe, using off-the-shelf optics and a 3D printed housing for high performance and low cost [25]. The spectrometer uses a 50 µm by 3 mm slit (National Aperture) at the entrance to isolate light from the image of the PLFA. An achromatic collimating lens (${f_1}$ = 75 mm) was used to collimate the light from the slit onto a diffraction grating (LightSmyth Technologies, 1500 lines/mm) which separates wavelength as a function of angle. A multi-component focusing lens consisting of two 2” achromats, $f$ = 250 mm, and a plano-convex lens, $f$ = 300 mm, was used to reimage this light onto the detector. The general approach was to place two identical achromatic lenses back-to-back in a symmetric configuration, eliminating odd-order aberrations such as coma, transverse color, and distortion. The plano-convex lens reduced spherical aberration from the other lenses in the system. A camera (Grasshopper3, FLIR, 5.86 µm pixels), with coverglass removed to prevent common-path interference, was used to detect the interference spectrum in each row as a function of angle across each column. The resulting system exhibits a pixel-limited spectral resolution ${\delta _s}\lambda $ of 33 pm, and a FWHM spectral resolution ${\delta _r}\lambda $ of 49 pm as measured with a calibration lamp. Each acquisition by the a/LCI system took 30 msec.

2.5 Data acquisition and analysis

A custom GUI for the clinical system was written using LabVIEW to control both a/LCI and OCT processing systems. Live displays of processed OCT B-scans and a/LCI scans were presented. A simplified live a/LCI scan was presented without background subtraction to preserve the frame rate and provide real time feedback for the user on signal fidelity. When diagnostic scans were acquired, integrated shutters were used to obtain background scans for the sample and reference fields to ensure proper background subtraction. This live mode enhanced a/LCI image acquisition speed in real-time for users, while improving signal fidelity during tissue scans.

2.5.1 Spectral-domain OCT processing

Standard spectral-domain OCT imaging procedures were applied for the conversion of raw interferometric data into 2D depth profiles (B-scans), involving background subtraction, k-space interpolation, and dispersion compensation [26]. Given the rotational nature of the OCT scan, the native B-scans derived from our images were in polar coordinates. To address the non-uniform rotational distortion (NURD) [27] due to variable angular velocity, a customized correction method was implemented as shown previously in [19]. Initially, eight pivotal positions were identified on the polar B-scan, coinciding with key points on the cross-sectional profile formed by the rectangular paddle cavity walls and tissue surface. Four central positions are used corresponding to the center of each side of the internal cavity, while the remaining four were the cavity’s corners. Subsequently, genuine angular positions of each A-line were estimated through spline interpolation, utilizing identified A-line positions as inputs and their known angular positions as outputs. This approach facilitated shifting the A-lines to align with a linear angular progression, followed by angular dimension interpolation to seamlessly bridge gaps in the B-scans. To finalize the process, we remapped polar coordinate images to rectilinear coordinates and cropped out portions of the image that did not include the sample.

2.5.2 a/LCI signal processing

To isolate the interferometric component of the detected spectrum, background subtraction was performed. Two inductive optical shutters (DACO Instruments, Milford, CT) with 25 ms switching times were controlled using a LabVIEW digital I/O card, one to block the sample arm and one to block the signal arm. For each diagnostic a/LCI scan, four exposures (total, sample only, reference only, and dark) were taken to permit isolation of only the interference term. The integration time of each image was 35 ms, with a 100 ms waiting period between exposures to allow the shutters to settle. After the dark, reference, and sample exposures, a burst acquisition of ten total scans (10 × 35 ms, or 350 ms total) was taken, and averaged after a/LCI processing to improve SNR. The total time to acquire ten scans and associated background images is less than 1 second.

a/LCI processing has been described extensively in previous publications [28]. Briefly, each a/LCI image consists of a 2D map of scattering intensity as a function of angle and wavelength (Fig. 5). Images are first background-subtracted and resampled to be linear in wavenumber. The result is then multiplied by a polynomial phase-correction term to compensate for dispersion, and Fourier-transformed line-by-line in the wavenumber dimension. Because the scattered field is discretized across an array of single-mode fibers, the “sample” background image is used to isolate only those pixels which contain information from the fibers. Since the fibers have a mode-field diameter slightly larger than the camera pixels, each discrete angle is identified as three rows of pixels centered on the peak intensity from each fiber. These lines are identified and binned to improved SNR. The result is a 2D map of scattering intensity as a function of scattering angle and depth within the sample (Fig. 5).

 

Fig. 5. (Left) a/LCI camera acquisitions showing total, sample, reference, and dark fields, with an inset showing interference from a few fibers. The four frames are used to isolate the interferometric term in the signal, using the formula ${|{{E_R} + {E_S}} |^2} - {|{{E_S}} |^2}\; - \; {|{{E_R}} |^2}\; + \; {|{{E_D}} |^2}$ (Right) a/LCI scans at various stages of processing, showing scattering intensity as a function of angle and sample depth. Since the sample field is discretized along single fibers, lines containing signal are selected prior to analysis (line selection). The A-line from each fiber must also be manually adjusted to account for pathlength variance (depth correction). In the final processed image, the coverslip is clearly visible. The measured thickness of the coverslip (assuming a refractive index of 1.51) was 149.8 µm, well within the 130-160 µm range characteristic of No. 1 coverglass.

Download Full Size | PDF

To validate the device, a/LCI scans of scattering phantoms were acquired. Phantoms were constructed using polystyrene microspheres of various sizes (6, 7, 8, and 10 µm from Polysciences, Inc., Warrington, PA) which were centrifuged, dried in a vacuum overnight, and combined with polydimethylsiloxane (PDMS) at a ratio of approximately 1 µL microspheres per mL PDMS. The mixture was sonicated to prevent clumping, poured into a mold, placed under vacuum to remove bubbles, and cured in an 80 °C oven for ∼1 hour.

Six a/LCI scans of each phantom were acquired, with repositioning after each scan to ensure nondegenerate measurements. From each background-subtracted, processed scan, an angular spectrum was acquired by binning over ∼460 µm in the depth dimension. After binning in depth, the resulting angular profile was normalized, and a second-order polynomial was subtracted to isolate the oscillatory component. A low-pass filter served to reduce speckle noise. The resulting spectrum was compared against a library of 131 Mie spectra, representing scattering by polystyrene microspheres (n = 1.58 at 830 nm) in PDMS (n = 1.41 at 830 nm) ranging from 5-18 µm in diameter in increments of 100 nm. The library spectra were also normalized and de-trended, and subsequently resampled to match the fiber sampling of the probe. The acquired angular scattering profile was compared with each profile in the library, and a chi-squared error was calculated. The profile with the lowest error was used to determine the scatterer size.

2.6 Clinical imaging

In an ongoing clinical pilot study conducted at the Center for Esophageal Diseases and Swallowing (CEDAS) at University of North Carolina, the clinical usability of the combined OCT and a/LCI device was evaluated. Human subjects research was approved by the Institutional Review Boards at Duke University Health System (Pro0090173) and the University of North Carolina (UNC; 17-3037). The study enrolled 31 patients, drawn from the cohort of BE patients under care at CEDAS. Each patient had 1-4 sites examined using the probe, followed by a physical biopsy and histopathological evaluation. The pathology diagnosis was compared with a/LCI measurements of nuclear size. Representative data for Barrett’s esophagus, more specifically low-grade dysplastic (LGD) esophageal tissue and non-dysplastic (NDBE) tissue, were selected for this report as a selected group of the data from the greater study to specifically validate the effectiveness of the combined OCT and a/LCI instrumentation. A larger clinical study is currently underway.

3. Results

3.1 Performance of the PLFA

Prior to analysis of the complete system, the PLFA was characterized for optical pathlength precision and fiber placement accuracy. Images of the polished distal surface were analyzed for fiber positioning in MATLAB and ImageJ. The slow axis of each PM fiber (along which polarization is maintained) was measured to be colinear with the array of fibers to an accuracy of 2°. The SM fibers were also analyzed for packing and even spacing. The active span of the 32 fiber elements (from PM core to PM core) was 2.456 mm, which is within 1% of the expected distance of 2.48 mm given the cladding diameter of 80 µm. The linearity of the fiber cores was also examined since the a/LCI software assumes 1D sampling of the scattered field with a constant angle increment Δθ. From a line drawn between both cores of the PM fibers, the maximum perpendicular deviation from this line by an SM fiber core was only 16 µm, representing an aspect ratio of the detection area of >150:1. The fiber spacing was also highly regular. The precise coordinate of each fiber core was identified in MATLAB, and mapped onto the expected angular range of the a/LCI device. An assumption of constant spacing between elements in our LFPA would result in the average detection channel misidentifying its corresponding scattering angle by <0.1°, with a worst case of <0.25°. Since Mie oscillations from cells occur over several degrees, this is well within the acceptable limit.

A sample a/LCI scan prior to pathlength correction is shown in Fig. 5 (top right). The variability in optical pathlength of each fiber element was determined by using the interference peak from the front surface of the coverslip glass. The total span of relative optical pathlengths between fiber elements was 228 µm, significantly less than the 5.2 mm imaging depth ($\frac{{\lambda _0^2}}{{4{\delta _s}\lambda }}$), or the 3.1 mm 6 dB rolloff point ($\frac{{ln(2 )}}{\pi }\frac{{\lambda _0^2}}{{{\delta _r}\lambda }}$) of the spectrometer [29]. Since fibers were matched in pathlength to a standard of 100 µm (±50 µm) it can be inferred that the assembly process introduced an additional pathlength variability of ±65 µm or so, likely attributable to slight movement of the fibers during the curing process.

3.2 Performance of a/LCI system

To characterize performance of the a/LCI system, scans from scattering phantoms were analyzed. Illustrative a/LCI scans are displayed in Fig. 6. The a/LCI analysis easily distinguished between similar 7 and 8 µm microsphere phantoms, and enabled accurate size determinations in line with their known scatterer sizes, confirming the miniature probe's proficiency in obtaining scattering data suitable for particle sizing. The angular range of the probe was quantified at approximately 2°-38°, marking a 10° extension from a previous design [23].z This extension, however, produced a reciprocal reduction in angular resolution, empirically determined to be $\Delta \theta = 1.27$°—still within acceptable clinical a/LCI device limits [30]. The system demonstrated remarkable signal strength across various depths, with observable scattering signal spanning up to a full millimeter in depth for most scattering angles and around 700 µm for high-angle scattering, which typically yields lower amplitude scattering. Angular scattering profiles (Fig. 6) were depth-binned (∼480 µm) for precise scattering peak identification during probe characterization. A calibration curve (Fig. 7) illustrates the miniature probe’s performance. Employing manual repositioning of the probe between scans, microspheres of 6, 7, 8, and 10 µm diameter were sized at 6.12 ± 0.18 µm, 6.96 ± 0.05 µm, 7.96 ± 0.05 µm, and 10.12 ± 0.12 µm, respectively, with a mere 96 nm mean error. This level of precision exceeds that expected from a diffraction-limited optical microscope, underscoring the miniature probe’s exceptional capacity for characterizing microscopic scattering objects in this size range. We note that the typical data in Fig. 6 were acquired prior to lens truncation. However, the data in the calibration curve in Fig. 7 were all acquired with the truncated lenses, showing that reliable sizing information is obtained.

 

Fig. 6. Representative scans using the miniature a/LCI probe without truncated lenses. a/LCI scans were taken using two very similar scatterers - 7.0 and 8.0 µm polystyrene microspheres in PDMS – to demonstrate sizing precision. The paddle accurately characterized the 7 µm spheres as 7.0 µm and the 8 µm spheres as 8.0 µm. All Mie oscillations are clearly visible and match their expected locations. The miniature probe exhibits an angular range of 2°-38°.

Download Full Size | PDF

 

Fig. 7. (Left) Calibration curve demonstrating sizing of microspheres in a/LCI. 6, 7, 8, and 10 µm microspheres were accurately identified as 6.03 ± 0.11 µm, 6.98 ± 0.07 µm, 7.98 ± 0.12 µm, and 10.15 ± 0.05 µm. Mean absolute error for all measurements is ∼0.1 µm, and all measurements precisely characterize microspheres to sub-wavelength accuracy.

Download Full Size | PDF

3.3 Application to In vivo imaging of Barrett's esophagus tissues

Illustrative clinical results from the integrated a/LCI and OCT platform are presented in Fig. 8. The upper segment features OCT images acquired from individuals diagnosed with Barrett's esophagus, including cases with and without dysplasia. Accompanying these OCT images are angle-resolved depth scans of light scattering from the tissue, presented in the middle. Furthermore, angular scattering distributions from the basal epithelial layer are displayed, alongside the corresponding best-fit Mie theory solutions (indicated by red lines) in the lower part of the figure. Overcoming challenges posed by factors like overlying mucus or air during a/LCI measurements is crucial for reliable interpretation [31]. This is managed by using OCT images as guides to the correct depth of tissue histological features. This depth guide facilitates the alignment of a/LCI measurements with the basal epithelial layer. In previous studies, we identified that the basal layer is most diagnostically useful, located at 200-300 beneath the tissue surface [11]. Notably, the combined instrument shows the potential to improve the diagnostic precision of a/LCI by harnessing the synergistic information afforded by OCT imaging. Noteworthy differentiation between dysplastic and non-dysplastic tissues is observed: Barrett's esophagus with low-grade dysplasia (BE w/ LGD) yielded an average nuclear size of 13.83 ± 0.66 µm while the NDBE data yielded an average nuclear size of 8.10 ± 0.50 µm. These measurements show a clear distinction between dysplastic and non-dysplastic tissue, following previously established benchmarks for a/LCI measurements of BE tissues [11,12,32].

 

Fig. 8. Data from the combined a/LCI and OCT platform in patients with Barrett's Esophagus, depicting (A) low-grade dysplasia and (B) non-dysplastic tissue. Top: OCT images from BE patients with and without dysplasia. Arrow shows trapped air between tissue and coverglass which can confound a/LCI measurements without image guidance. Scale bar is 1 mm in tissue. Center: Angle-resolved depth scan of light scattered from tissue. Red boxed area shows depth selected for analysis, 200-300 µm beneath the tissue surface. Bottom: Example angular scans for 2 tissue types (blue line) with best-fit Mie theory solutions (red line). Analysis of these angular scans produces a nuclear size of 13.8 µm for the LGD tissue and 8.1 µm for the non-dysplastic tissue.

Download Full Size | PDF

4. Discussion

The ability of a/LCI to measure depth-resolved nuclear morphology has shown promise for diagnosing dysplastic tissues. Recent work in detecting cervical dysplasia with a/LCI has produced optimized instrumentation [33] and machine learning-based analysis methods [34] to improve clinical utility. Here we suggest a new form factor for esophageal imaging that will likewise improve clinical utility. These advances have focused on a new form factor for the optical probe, incorporation of OCT imaging for alignment with tissue features, and a lower cost optical engine for both a/LCI and OCT modalities.

One of the primary advantages of the new a/LCI probe is the use of single-mode fibers compared to the leached fiber bundle approach used previously [17]. The single-mode fibers do not present incoherent noise from non-fundamental modes and exhibit substantially higher throughput (>99.93%/meter) compared to the leached imaging bundles (10-20%/meter). As a point of comparison, the a/LCI scans shown here contain observable fringes from polystyrene scattering phantoms up to 700-1000 µm in optical depth (Fig. 5) compared with the 500-700 µm seen in systems using a leached imaging bundle [17]. It is likely that increased signal from the SM fibers contributes to improved depth imaging.

The key diagnostic advantage of a/LCI is its ability to provide depth-resolved measurements. In most epithelial tissues, dysplasia originates at the basal layer, located roughly 250 µm below the tissue surface [8,12]. By traditional endoscopy with histological sampling, the presence of dysplasia is only detectable when it appears on the mucosal surface in sufficient quantity to be recognized. By this time, it is possible that the window for minimally invasive intervention has closed, and frank invasive cancer is present. Our probe was designed for optimal focus in the basal layer where neoplasia originates, with the focal depth occurring approximately 285 µm past the surface of the coverslip.

In a/LCI, angular resolution is characterized not only by fiber spacing, but also by the error in collimation of the illumination beam, since the angular spectrum will be convolved with the distribution of angles incident on the sample. In Zemax, the lens spacings for the probe body were optimized until a minimum collimation error of ±0.43° was achieved. Since this is less than our angular sampling of approximately 1.27°/fiber, the measured angular spectra are not significantly contaminated by this effect. Empirically, excellent collimation was observed over 1-2 meters on the benchtop, despite the use of gradient-index optics normally susceptible to aberration. In a previous study, it was shown that minimal degradation in diagnostic performance is achieved so long as Δθ is less than ∼3.6° [30]. Future work could involve a reduction in the number of fibers in the PLFA to around ten, substantially reducing the time required to build a probe. Alternatively, distal optics with a larger numerical aperture and magnification could be introduced, sacrificing angular sampling for a wider angular range. The use of custom optics may enable a reduction in the outer diameter of the probe body, improving clinical utility.

One of the primary practical benefits of this system over previous a/LCI instruments is the cost. Because each single-mode fiber only costs ∼$\$$6/meter, a 30-fiber probe will cost $\$$180/meter, or ∼$\$$220/meter including the cost of both PM fibers. The total cost of distal optics (three lenses and prism) in each probe is ∼$\$$350, costing a total of ∼$\$$1000 for a 2.5-meter probe. The costs of 3D-printed parts, epoxy, and polyolefin tubing are almost negligible. Our probe is substantially cheaper than previous designs, which incurred costs of >$\$$1,000/meter in components alone. Additional cost savings were realized in the newly designed imaging spectrometer. Older a/LCI instruments [17] used a commercially available spectrometer, coupled with an expensive scientific CCD detector with a low frame rate (<10 Hz). In this system, we used a CMOS detector with a similar quantum efficiency, capable of 163 frames per second, whose total cost is under $\$$1,000. Use of multi-element lenses in the imaging spectrometer enable a high spectral resolution of 33 pm, more than doubling the imaging depth and allowing for a wider variation in optical pathlength of the SM fibers. A custom 3D-printed housing for the spectrometer components further serves to reduce cost. One disadvantage of the PLFA is the time required to path match individual fibers. While a skilled operator could cut and match up to 3-4 fibers/hour, an automated process to actively polish and monitor each fiber for optical pathlength would substantially improve this technique.

Another significant advance in this probe and system is the integration of co-registered a/LCI measurements with real-time OCT imaging guidance. In previous clinical studies, a high rejection rate of a/LCI scans was seen of up to 55% [12]. This arises from low signal intensity, mainly because of probe movement, insufficient contact of the probe with the tissue and poor tissue appositions. In this study, the visual guidance of OCT helped orient a/LCI measurements by ensuring good tissue apposition such that layered structures are even across the field of view and depth resolved scattering information can be retrieved. Previous studies have also shown that overlying mucus can be a confounding influence when trying to determine the start of the tissue directly from the a/LCI measurements [12]. The OCT tomograms provided useful tissue guidance, as they can help the clinician avoid acquiring poor scans where the presence of air or mucous. Further the OCT tomograms permit a clear registration of a/LCI measurements at the basal layer of the epithelial tissue.

An initial demonstration has been presented involving patients with Barrett's esophagus (BE), utilizing the paddle-based dual-modality probe. The primary objective of this work has been to fully describe and characterize the new probe, and to showcase early clinical results differentiating non-dysplastic and dysplastic tissues. The single sample from each tissue type was consistent with results from earlier studies of BE using a/LCI [12]. The non-dysplastic tissue sample showed a mean nuclear size in the basal layer of 8.1 mm while the sample from the low grade dysplasia tissue showed a mean nuclear size in the basal layer of 13.83 mm. These nuclear sizes align well with the pre-determined decision line of 11.84 mm with the non-dysplastic tissue sample falling below this threshold and the dysplastic tissue sample exceeding it. A larger, multi-year clinical effort is currently underway.

5. Conclusions

In this work, the comprehensive development and clinical implementation of a dual-modality optical imaging system designed to capture a/LCI diagnostic scans and OCT tomograms of esophageal epithelial tissue during routine upper endoscopy has been presented. The emphasis has been placed on improving the design of the a/LCI optical probe while lowering cost and seamlessly integrating it with a rotational endoscopic OCT probe within a slim external paddle, affixed to the endoscope. The new a/LCI probe uses a new design which incorporates a linear array of single mode fibers and enables side viewing. The former advance helps with manufacturing cost while the latter aids in clinical utility. The incorporation of OCT imaging into the same probe enables visual guidance and orientation of a/LCI measurements while avoiding lost data due to poor tissue apposition. Clinical application of the probe was demonstrated by obtaining B-scans of both dysplastic and non-dysplastic Barrett's Esophagus (BE) esophageal epithelium using OCT and a/LCI measurements of nuclear morphology which aligned with previous studies. The clinical utility of the device will be defined in an investigation to evaluate its potential in prospective detection of dysplasia within the esophageal epithelium.