Open Access
Add to BibSonomy
Abstract
We present a prospective clinical study using angle-resolved low-coherence interferometry (a/LCI) to detect cervical dysplasia via depth resolved nuclear morphology measurements. The study, performed at the Jacobi Medical Center, compares 80 a/LCI optical biopsies taken from 20 women with histopathological tissue diagnosis of co-registered physical biopsies. A novel instrument was used for this study that enables 2D scanning across the cervix without repositioning the probe. The main study goal was to compare performance with a previous clinical a/LCI point-probe instrument [Int. J. Cancer 140, 1447 (2017) [CrossRef] ] and use the same diagnostic criteria as in that study. Tissue was classified in two schemes: non-dysplastic vs. dysplastic and low-risk vs. high-risk, with the latter classification aligned with clinically actionable diagnosis. High sensitivity (non-dysplastic vs. dysplastic: 0.903, low-risk vs. high-risk: 1.000) and NPV (0.930 and 1.000 respectively) were obtained when using the previously established decision boundaries, showing the success of the scanning a/LCI instrument and reinforcing the clinical viability of a/LCI in disease detection.
© 2020 Optical Society of America under the terms of the OSA Open Access Publishing Agreement
1. Introduction
Cervical cancer is one of the most common forms of cancer worldwide, with approximately 570,000 new cases and 311,000 deaths occurring in 2018 [1]. Although improved detection of precancerous lesions (dysplasia), using cytology based methods such as the Papanicolaou test (Pap smear), the current gold standard, has led to a significant decrease in incidence and mortality, the recent flattening of these trends suggest that this approach may have reached its limit [2]. During routine cervical screening, abnormal cytology from a Pap smear is followed by colposcopic examination of the cervix to identify dysplasia, but cytology and colposcopy suffer from low sensitivity [3–5] which is likely related to the subjective nature of both methods. More recently, introduction of vaccination against infection by the human papillomavirus (HPV) is projected to substantially reduce the number of cervical cancer cases over the next forty years [6], but logistical issues with roll-out, especially in low-income countries where cervical cancer is disproportionately more prevalent, mean that there is currently still an unmet need for a high-accuracy quantitative screening method of cervical dysplasia.
Persistent infection from the human papillomavirus is responsible for virtually all cases of cervical cancer [7,8]. HPV infection can cause noticeable changes in the morphology of the cervical epithelium and the resulting cervical intraepithelial neoplasia (CIN) is classified depending on the severity of the morphological changes [9]. Low-grade squamous intraepithelial lesions (LSIL), also referred to as CIN-1, are characterized by an increase in nuclear size, irregularity of shape, compactness, and other morphological factors, all generally restricted to roughly the lower third of the epithelium just above the basal lamina [10]. High-grade squamous intraepithelial lesions (HSIL), also classified as CIN-2 or CIN-3, present cytonuclear atypia in the form of increased nuclear-to-cytoplasmic ratios as well as increased severity of previously mentioned nuclear abnormalities that become pervasive in increasingly superficial layers of the epithelium [10].
In this paper, we utilize angle-resolved low coherence interferometry (a/LCI) for identifying dysplasia in cervical epithelial tissue in vivo. A variety of different optical techniques have been previously applied to this problem, including optical coherence tomography [11], confocal microscopy [12], Raman spectroscopy [13], and fluorescence and reflectance imaging [14]; a/LCI is unique in that it can provide quantitative information about the nuclear morphology of the cervical epithelium and do so with depth resolution to isolate measurements of the tissue at the basal layer of the epithelium where morphological deviations resulting from dysplasia begin. We have previously demonstrated the feasibility of a/LCI in distinguishing between benign cervical epithelial tissue, low-grade dysplasia, and high-grade dysplasia with high accuracy [15]. However, the point-probe nature of the previous instrument was a significant limitation because of the inability to evaluate an appreciable region of the cervix. Here, we solve this problem by presenting a scanning system with electronically controlled optical components that allow examination across an 8 mm diameter field of view without needing to reposition the probe.
The goal of the clinical study is to assess the diagnostic capabilities of scanning a/LCI for detecting cervical dysplasia in vivo. We apply the decision line established in our prior a/LCI study [15] to prospectively classify cervical dysplasia and analyze the accuracy of this approach. We also compare the performance of the newly developed compact handheld scanning a/LCI instrument with that of the previously published point-probe instrument study to establish the feasibility of wide area tissue scanning with a/LCI, an important requirement for future clinical adoption.
2. a/LCI Instrumentation
2.1 Introduction to a/LCI
a/LCI quantitatively assesses nuclear morphology by analyzing the angular backscattering of light from the cervical epithelium. The core instrumentation for a/LCI is well-established and consistent across previous implementations as shown in Fig. 1. [15–21]. Light from a superluminescent diode source is split between a reference arm and a p-polarized sample arm which feeds to a specialized fiber delivery probe. The probe consists of a polarization-maintaining (PM) fiber which outputs collimated light at an oblique angle to the tissue surface and a coherent fiber bundle to collect the elastically backscattered light as a function of scattering angle by exploiting a GRIN lens to produce a Fourier transform. The light conveyed by the bundle is then recombined with the reference field to form an interferogram on an imaging spectrograph, with each channel corresponding to a scattering angle and detected individually with a CCD camera. To extract information about nuclear morphology, we perform Mie theory based inverse light scattering analysis (ILSA) [22]. In ILSA, theoretical angular scattering profiles are generated for a range of scatterer sizes and relative refractive indexes to populate a library, which is used to iteratively compare to the sample angular scattering profile from tissue until a best fit is found using a chi-squared analysis. Thus, two aspects of nuclear morphology are extracted: the size of the nucleus, and the ratio of nucleus refractive index relative to that of the cytoplasm, also known as the nuclear density. The use of interferometry in a/LCI allows for simultaneous acquisition of morphological information at a range of depths. Not only does this allow a/LCI to analyze tissue below the most superficial layers, but also to selectively examine depths of interest.
Fig. 1. A schematic of the core instrumentation for a/LCI, with Mach-Zehnder interferometer geometry. The probe tip consists of sample illumination via a PM fiber and backscattered light collection with a coherent fiber bundle [16].
a/LCI is typically a contact imaging technique in clinical applications, so design of the probe end of a/LCI instrumentation requires special considerations for different areas of the body. The initial clinical instrumentation for cervical analysis had a probe tip which funneled the specialized fiber delivery probe through a handheld wand consisting of a stainless steel tube with a bend incorporated at the end to conform to the anatomy of the vaginal canal, as well as a GRIN lens that served to both deliver collimated, off-axis light to the sample from the PM fiber and map the backscattered light to the fiber bundle as a function of angle. This probe design had the advantages of being robust and easy to manufacture due to its simplicity, but had limited clinical utility due to the difficulty of reliably repositioning the probe at different positions on the cervix, which is essential for clinical utility, especially if the clinician sought to examine a specific region that is suspected to harbor dysplasia. The contact probe nature of the instrument means that implementing optical elements to remotely manipulate a beam must either be compact enough to fit inside the limited space of the probe housing or occur upstream in the optical train such that they are always outside of the body even if the probe is fully inserted. Additional difficulty comes from the fact that in a/LCI, the collected light does not follow the same path as the incident beam, and the range of angles at which light is collected maps to different points on the objective lens plane, requiring an aperture that is especially large considering the desire for a handheld, compact probe to easily access the cervical epithelium [21].
A scanning system for a/LCI was initially presented as a benchtop system by Steelman et al. [17]. In a/LCI, mapping of scattering angles to specific elements of the collection fiber bundle means that the light collected by the slit of the imaging spectrometer is a one dimensional scattering distribution, which we term the collection axis. With scanning a/LCI, scanning is performed perpendicularly to the collection axis to take advantage of the remaining clear aperture on the lens plane, and this direction is termed the scanning axis. The scanning mechanism is implemented using a reflection-only three-optic rotator (ROTOR), which is an image rotator analogous to a Dove prism, that rotates the scanning and collection axes together. The radial positioning from the ROTOR is combined with lateral scanning using a gimbal-mounted two-axis mirror to achieve 2D scanning. More detailed information about the scanning system can be found in Steelman et al. [17] The result is an asterisk-shaped scanning profile which allows any point within a circular field of view to be measured. We developed a compact, portable version of the ROTOR to incorporate into a handheld probe for the present study (Fig. 2).
Fig. 2. Optical design for the miniaturized ROTOR. A GRIN lens G1 is butt-coupled to the a/LCI probe lens G0 and 3 mirrors RM1-RM3 form the reflective surfaces. Two identical lenses RL1 form a 4f system that prevents the beam from clipping off of RM2 and a final GRIN lens G2 is used to collimate the fields. The distinctly colored ray paths correspond to optical fields of different angles.
2.2 Optical design
Scanning a/LCI capabilities are implemented as an extension of the original a/LCI engine by inserting the fiber bundle from the a/LCI engine and passing the light through the scanning elements of the new probe. The full optical design for the scanning probe tip is shown in Fig. 3. Optimization of the compact ROTOR was important to allow the use of 1/2”-diameter optics. Two GRIN lenses, G1 and G2 were used, one at the input of the scanner, after the GRIN lens G0 of the original a/LCI fiber probe, and the second at the exit of the scanner to preserve the numerical aperture of the beam of light leaving the ROTOR. , The 4f system formed by identical achromatic triplet lenses RL1 creates an image of both the phase and amplitude of the field while also ensuring the beam diameter of the field is contained within the diameter of RM2. To house these optical components, a custom mount is 3D-printed with pre-aligned mounting slots for the two GRIN lenses and two identical triplet lenses RL1. To account for material tolerances, the mirrors RM1 and RM3 are attached separately via tip-tilt and rotation stages so that the ROTOR can be properly aligned, as seen in Fig. 4(b).
Fig. 3. Visualization of probe body showing spatial orientation of components. The a/LCI probe tip is the same one as in Fig. 1, with a PM fiber for propagating the excitation beam and a fiber bundle for collecting and propagating scattered light from the sample to the a/LCI system..
Fig. 4. Physical depictions of the probe components. (A) is a side profile of the actual probe with detachable probe tip. (B) is a 3D transparent rendering of the ROTOR.
To allow for image rotation using the ROTOR, an integrated pulley is connected to the shaft of a stepper motor via a timing belt, and bearings are attached to each end of the ROTOR for smooth, free rotation. Additionally, limit switches are attached to the walls of the handheld probe to provide a homing mechanism for reproducible orientation of the ROTOR during data acquisition. The stepper motor driver and limit switches are controlled by a custom circuit board attached which is interfaced using an Arduino based microcontroller.
2.3 2D scanning
To implement automated scanning, a commercial voice coil mirror (MR-15-30, Optotune) was used. This mirror was chosen for desirable properties including a small form factor, high mechanical tilt angle, and near-gimbal rotation. An aspheric lens L1 with high numerical aperture (NA = 0.50) and short focal length (f = 20 mm) was placed after the final ROTOR GRIN lens to minimize spherical aberrations while confining the optical fields to the diameter of the fold mirror M1 and voice coil mirror. A Hastings triplet L2 was placed after the scanning mirror in order to refocus the fields while collimating the off-axis rays and reducing spherical aberrations. Two doublet lenses (L3, L4) relay the fields to the end of the probe, arranged to minimize the axial focus shift between paraxial and off-axis rays. A doublet (L5) lens and double-concave (L6) lens are further used to optimize field collimation and minimization of spherical aberrations. A curved meniscus lens L7 was placed at the distal end to achieve contact imaging to match the curvature of the cervix while broadening the radius of the lens aperture used to maximize the angular collection range of the device. Additionally, two wedge prisms (P1, P2) are incorporated to create an 8-degree bend in the probe to allow it to conform to the anatomy of the vaginal canal.
2.4 Mechanical design
The main body of the probe was 3D-printed using a standard rigid photopolymer (VeroWhitePlus, Stratasys), while the probe tip is made using a medical grade photopolymer (MED610, Stratasys), as shown in Fig. 4(a) and Fig. 4(b). The probe tip was designed in two halves with slot-in cavities for the optics for simple alignment. After lens placement, the halves were sealed by applying photopolymer to the seams followed by UV curing. To provide visual guidance to the clinical during insertion and positioning of the probe, a miniature USB color camera (MD-B1001, Misumi) was integrated into the probe tip along with a white light illumination LED. To allow for easy sanitation of the probe tip between uses, the probe tip was designed to be detachable, with a threaded twist lock used to reattach it to the main body, and 6-pin electrical connectors (Nano360, Omnetics Connector Corporation) to provide electrical connections to the camera and white light illumination LED. To hold the probe during disinfection, a custom threaded attachment was designed and 3D-printed and to enable use with a peroxide-based high-level disinfection unit (Trophon, Nanosonics). To ensure that fluid did not enter the internal portion of the probe, silicone O-rings were employed to seal the optical path and electronic connections between the probe and probe tip and prevent damage during disinfection. The fiber bundle from the a/LCI engine was mounted within the main probe body via custom printed fittings on an X-Y translation mount for alignment, incorporating a threaded twist lock and mount for strain relief of the fiber. All of the optical and electronic components were placed on a rolling multi-shelf cart with a small computer and touchscreen monitor on the top shelf.
2.5 Software
To control the instrumentation, custom software was written in LabVIEW to interface with all electronic components, including the voice coil mirror driver, stepper motor driver, a/LCI engine, limit switches, camera, and white light LED. The software, designed to permit ease of use in the clinic, first displays and records the live white light image from the camera for probe alignment, and then is followed by automated repositioning of the ROTOR and voice coil mirror for scanning. 36 a/LCI mapping scan points, whose positions can be seen in Fig. 5(b), are designated for creating a representative map of the FOV of the instrument by looking at signal intensity at each point. These are followed by twenty diagnostic a/LCI scans at each of the four quadrants, corresponding to 12, 3, 6, and 9 o’clock positions, which are processed to generate the nuclear morphology information used to assess a given site. The 36 mapping a/LCI A-scans were natively aligned in polar coordinates for each depth, then converted to Cartesian coordinates and interpolated into a 50x50-pixel volumetric image stack.
Fig. 5. Representations of the phantom used for scanning a/LCI validation. (A) A volumetric rendering of the four-quadrant phantom, showing location of 8 µm and 15 µm diameter polystyrene microspheres. (B) A map of the a/LCI scanpoints with an overlayed representation of the phantom quadrants in gray. (C) A white light image from the probe tip with the field of view of the instrument outlined in red. (D) A volumetric rendering of the signal intensity at the given a/LCI scanpoints [23].
3. Phantom validation
To validate the scanning capabilities of the instrument, tissue phantoms containing polystyrene microspheres of varying sizes embedded in PDMS were used (Fig. 5(a)). A custom mold was 3D printed using acrylonitrile butadiene styrene (ABS) with four quadrants to provide a model of clinical data collection at four quadrants in the cervix. The surface was contoured to match the curvature of the final lens L7 (Fig. 3). Two opposite quadrants were embedded with 8 um microspheres while the other two were embedded with 15 um microspheres, representing the general range of expected nuclear diameters in the clinical study. A 36-point a/LCI scan pattern was created over of an 8 mm diameter circular field of view to generate a representative map of the FOV of the instrument (Fig. 5(b)).The LabVIEW software was used to collect a/LCI scans to map the scattering from the phantom. The white light camera image and volumetric image of the phantom are shown in Fig. 5(c) and 5(d), with the support features within the mold clearly visible in both images along with the distinct quadrants. Twenty diagnostic a/LCI scans were acquired at each quadrant and processed using Mie-theory based ILSA to produce a best fit scattering profile and corresponding nuclear diameter which can be seen in Fig. 6, with good agreement between the measured and actual diameter values.
Fig. 6. Averaged a/LCI scans for each quadrant of the phantom along with corresponding a/LCI angular scattering and best fit profiles predicted from Mie theory-based ILSA [23].
Analysis of the phantom measurements show the scanning a/LCI instrument’s capability for 2D repositioning of the beam, as well as accurate and reproducible diameter measurements at four distinct quadrants. The system was able to size the microspheres in each quadrant with sub-wavelength accuracy (0.25 µm average error) with an angular collection range of $20.8^{\circ }$.
4. Clinical study
A clinical study to assess the ability of scanning a/LCI to prospectively grade tissue health was performed at the Jacobi Medical center in the Bronx, NY and was approved by the IRB of the Albert Einstein College of Medicine with concordance by the Duke University Medical Center IRB. Women above the age of 21 who had not had sexual intercourse for at least 24 hours prior to the study visit were recruited following informed consent. Exclusion criteria included positive urine pregnancy test, current gynecologic infections or discharge, and no prior surgery performed on the cervix.
In preparation for each patient, the probe tip was detached and disinfected by a trophon unit, and the clinical software was loaded. A speculum was inserted prior to the a/LCI probe; this was not necessary to use the scanning a/LCI instrument but allowed the clinician to collect subsequent tissue biopsies more easily. The clinician guided the probe to the cervix using the white light video from the camera at the distal end of the probe as a visual aid. Further this guidance could be used to ensure good contact was made onto the cervix (Fig. 7). After centering the probe over the transformation zone of the cervix, the a/LCI imaging sequence was initiated in the software, comprising 36 scan points spaced across an area of 50 mm$^2$ followed by 80 a/LCI biopsy scans taken at the four selected biopsy sites (20 per quadrant). At completion of the imaging sequence, the probe was removed and the speculum was left in place. A Wallach colposcope was then used to obtain a digital image of the cervix for co-registration of the transformation zone with the a/LCI image. Biopsy sites were selected based on colposcopic appearance, looking for sites that exemplified dysplasia. Biopsy specimens varied in size but typically range from 3-5 mm, with the biopsy site within a quadrant based on the area that appears most abnormal under the colposcopic magnification. Dysplastic areas were brought out by use of acetic acid and are described as acetowhitening, punctations, or mosaicism. Some patients underwent a previously scheduled loop electrosurgical excision procedure (LEEP) immediately following the data collection; tissue biopsies were taken either with the assistance of the colposcope or during the LEEP and all specimens were analyzed by pathologists to provide a histological diagnosis for comparison with the a/LCI measurements. Tissue biopsies for each quadrant were assigned a classification (negative, CIN-1, CIN-2, or CIN-3) as well as notes for any additional findings (koilocytic atypia, abnormal acetowhite epithelium, etc.)
Fig. 7. (A) Representative white light image from the probe tip showing transformation zone of the cervix as well as the locations of the 36 a/LCI scanpoints in blue. (B) Representative volumetric rendering of cervix at a/LCI scanpoints in (A). Labeled points correspond to the location of the 4 a/LCI biopsy sites and the averaged nuclear diameter at each quadrant. The cervical os is visualized as a dark spot just under the 12 o’clock data point. (C) Colposcopic image with physician-annotated os outlined in yellow, transformation zone in red, and edge of ectocervix in blue.
Women enrolled in the study had 36 a/LCI scans taken at the scanpoints as shown in Fig. 7(A) to perform volumetric imaging of the cervical epithelium. This was followed by 20 repeated a/LCI scans at the 12, 3, 6, and 9 o’clock positions corresponding to the location of physical tissue biopsies for a total of 80 a/LCI scans for classification of tissue co-registered with pathology.
One significant limitation of our study was the number of instrument malfunctions that occurred; out of the total of 41 women were enrolled in the study, data were only successfully acquired from 20 patients. These failures included mechanical breaks in the probe, electrical component failures, and software errors. Though the low yield in imaging is a significant limitation of this study, we implemented several design improvements to address these issues as they occurred, which will help ensure more robust performance in the clinical environment for future iterations of this device.
Each raw a/LCI A-scan was processed using a custom MATLAB script. An A-scan was created by summing the signal intensity across scattering angles at each depth to assess signal quality and orient the measurements. Data were discarded if the peak signal strength of this aggregate measurement did not reach a predefined power threshold ($<10^7$ counts). The start point of each depth profile was normalized to the surface of the cervical epithelium determined by a characteristic peak in a plot of signal vs. depth of the scan. The depth information was binned in 50 µm increments down to 500 µm below the surface of the cervical epithelium, and standard Mie theory ILSA was performed to obtain nuclear morphology information from each a/LCI scan at each depth bin. The depth bin corresponding to 200-250 µm below the surface empirically corresponds to the basal layer of the epithelium [15,18], so the nuclear diameter and density information from the 20 a/LCI scans taken at this depth were averaged to obtain a measurement of the average nuclear diameter and density at that given biopsy site. Representative raw a/LCI scans and intensity vs. angle plots can be seen in Fig. 8.
Fig. 8. (A) Representative depiction of patient raw a/LCI scan and (B) angle-resolved intensity profile. The pattern in (B) corresponded to a nuclear diameter value of 11 microns and a nuclear density of 1.073.
Non-dysplastic tissue presents differently from dysplastic tissue when processed biopsy specimens are viewed as histopathology slides; the classifications of low- and high-grade dysplasia are based on the location and severity of visual cues observed by the pathologist which can be subjective. However, the clinical impact of these two diagnoses is quite different, with low-grade dysplasia usually resolving without treatment such that a watch-and-wait approach is employed following diagnosis. On the other hand, the treatment for high-grade dysplasia is much more immediate and urgent [24]. Therefore, the data from the study is first dichotomized based on histopathological classification, as benign (non-dysplastic) tissue vs. LSIL + HSIL (dysplastic) tissue, and second according to the clinical treatment, as benign tissue + LSIL (“low-risk”) vs. HSIL (“high-risk”). For each of these two schemes data are plotted as a map of nuclear diameter vs. density and separated by a classification line. Each plotted point is color and symbol colored to indicate the pathological diagnosis. Comparisons of these groups were conducted using one-sided Student’s $t$-test considering the well-established trends in nuclear morphology.
Diagnosis by a/LCI measurements was made using a decision line established by the previous point probe a/LCI study, which was dependent only on nuclear diameter (decision line at 10.5 µm) to distinguish between non-dysplastic and dysplastic tissue [15]. The current, scanning a/LCI study uses this classification line prospectively. Performance was assessed first by calculating sensitivity and specificity based on the decision line and then by verifying that this is the optimal decision point by using a receiver operator characteristic (ROC) curve. For classification according to clinical treatment approach, a decision line that is a factor of both nuclear diameter and density was established using linear discriminant analysis (LDA), weighted to maximize sensitivity. This effect was seen in our previous study [15] and can be viewed as evidence that nuclear density becomes more of a predictive biomarker with increased severity of dysplasia [24]. For this analysis we also prospectively utilized the decision line previously determined by Ho et al. [15] and also compared its accuracy to one derived retrospectively using our new data and LDA. Corresponding ROC curves were also generated to establish the goodness of separation between low-risk and high-risk tissue based on a/LCI measurements.
5. Results
A total of 80 distinct optical biopsies (benign, n = 49, LSIL, n = 19, HSIL, n = 12) were collected from twenty women, each of whom received four co-registered physical biopsies at 12, 3, 6, and 9 o’clock quadrants relative to the transformation zone of the cervix. The signal quality of the a/LCI scans was consistent across quadrants for the same patient. Mie theory based ILSA was performed on all of the a/LCI data to obtain a measurement of nuclear diameter and density for each biopsy site. A summary of the results can be seen in Table 1.
Table 1. a/LCI optical biopsy results comparing scanning a/LCI instrument performance with point-probe instrument performance. Point-probe data adapted from [15]
5.1 Histopathology-based classification
The difference between non-dysplastic and dysplastic tissue is considered here (Fig. 9), with diagnoses of LSIL and HSIL classified as positive results and non-dysplastic tissue (or other benign reported conditions) classified as a negative result. The mean nuclear diameter at the basal epithelial depth bin at 200-250 µm below the surface was significantly higher for dysplastic tissue than for non-dysplastic (9.85 µm vs. 11.86 µm, $p < 0.001$). In the previous a/LCI study [15], the nuclear diameters for the same classification scheme at the same depth were also statistically significant with a higher mean diameter for dysplastic tissue (8.22 µm vs. 11.62 µm, $p < 0.001$). The mean nuclear density at the same depth was observed to be slightly lower for dysplastic tissue than non-dysplastic tissue, but this difference was not statistically significant (1.053 vs. 1.050, $p > 0.05$). The mean nuclear density from the previous a/LCI study was significantly lower in dysplastic than non-dysplastic tissue (1.053 vs. 1.042, $p < 0.001$).
A scatter plot of the data using the histology-based classification scheme is presented in Fig. 10(a). The previous study established an optimal decision line of 10.5 µm between non-dysplastic and dysplastic tissue, independent of nuclear density. Applying the line to this study demonstrates high sensitivity and specificity (sensitivity = 0.903, specificity = 0.816) in distinguishing between dysplastic and non-dysplastic tissue. To evaluate the diagnostic efficacy in the current study, a receiver operator characteristic (ROC) curve was plotted using nuclear diameter only to separate the data. This analysis produces an area under the curve (AUC) of 0.883 (Fig. 10(b)). The optimal decision line based on the ROC curve was found to be 10.47 µm, in close agreement with the prospective decision line of 10.5 µm. A negative predictive value (NPV) of 0.93 and positive predictive value (PPV) of 0.756 were obtained for this classification set. The overall accuracy was 0.85, compared to 0.98 in the point-probe study.
Fig. 9. Comparison of nuclear diameter and density information between non-dysplastic and dysplastic patients.
Fig. 10. (A) Scatter plot of nuclear diameter and density values computed from Mie theory-based ILSA, with a prospective decision line at 10.5 µm shown in blue. (B) ROC curve for histology-based classification with the prospective decision line represented in blue
5.2 Treatment-based classification
We next considered the treatment-based distinction between low-risk and high-risk tissue (Fig. 11), with HSIL considered a positive result, and benign and LSIL considered negative results. At the basal layer of the epithelium, the average nuclear diameter was higher for high-risk tissue than for low-risk tissue, and the difference was statistically significant (10.36 µm vs. 12.16 µm, $p<0.001$). This once again resembles the results from the previous study which reported a statistically significant higher average mean diameter for high-risk tissue (9.27 µm vs 12.04 µm, $p<0.001$). The nuclear density values for high-risk tissue was lower than for low-risk tissue, and the difference was statistically significant (1.053 vs. 1.045, p = 0.03). The previous study also reported lower values for nuclear densities in high-risk tissue vs. low-risk tissue, and the differences were statistically significant (1.050 vs. 1.041, $p < 0.001$).
The same scatter plot information is separated by each classification of tissue in Fig. 12(a). Because nuclear density had more of an influence on the results when distinguishing between low- and high-risk tissue, a line as a function of both nuclear diameter and density was generated to separate low- and high-risk tissue using linear discriminant analysis (LDA) weighted to maximize sensitivity. Using this line, a sensitivity of 1.000 and specificity of 0.706 were achieved with an overall accuracy of 0.750. The AUC for the ROC plotted using this sloped line was 0.846, showing generally good separation between low- and high-risk tissue (Fig. 12(b)). The NPV and PPV were 1.00 and 0.375, respectively.
Fig. 11. Comparison of nuclear diameter and density information between low-risk and high-risk patients.
Fig. 12. (A) Scatter plot of nuclear diameter and density values computed from Mie theory-based ILSA, with classification line shown based on linear discriminant analysis. (B) ROC curve for treatment-based classification with the prospective decision line represented in blue.
6. Discussion
Classification of dysplasia using the prospective decision line at 10.5 µm produces high sensitivity, specificity, and accuracy between quadrants, demonstrating the instrument’s capability to distinguish between dysplastic and non-dysplastic tissue at multiple points across the cervix. The fact that the prospective threshold very closely matches the optimal threshold obtained via ROC analysis of the scanning a/LCI data indicates that using a quantitative threshold based on a/LCI nuclear morphology for classification, is a robust surrogate biomarker of dysplasia., an important criteria for its viability as a clinical diagnostic. The treatment-based classification scheme showed perfect sensitivity and negative predictive value, a valuable characteristic for a clinical screening tool, as a follow-up colposcopy may be avoided if the result from the instrument is negative. Specificity, positive predictive value and overall accuracy were not as high in this study as in the previous retrospective study, but this is consistent with prior a/LCI results due to the lack of clear morphological distinction between LSIL and HSIL [15]. The decision line obtained here by LDA is similar to that from the previous study, but more work is needed to be able to better define a decision line for treatment-based classification in future studies.
The clinical scanning a/LCI instrument measured a statistically significant increase in nuclear diameter for non-dysplastic vs. dysplastic as well as for low-risk vs. high-risk tissue classifications at the basal layer of the cervical epithelium, corroborating results from previous clinical a/LCI results [15,18]. The diameter values for both dysplastic tissue and high-risk tissue obtained here are very similar to those from the previous a/LCI clinical study, suggesting that it is feasible to assign quantitative diameter ranges to specific stages of dysplasia in the cervix. However, the values of the nuclear diameters for non-dysplastic or low-risk tissue are somewhat different when compared with the previous study, with the average values being over 1 µm higher in both classification schemes. This may be attributable to the present cohort being recruited from a clinic consisting primarily of patients with cervical dysplasia, with every imaged patient presenting with dysplasia present in at least one quadrant of their cervix or with koilocytic atypia, in contrast with the previous study, which included a cohort of healthy controls. This seems to imply that morphological changes in the cervix may be more ubiquitous than previously thought, possibly due to the field effect of carcinogenesis [25,26]. This too may explain the lower accuracy and false positive rate when using the scanning a/LCI instrument vs. the previous point-probe instrument since accuracy is determined in this study by quadrant instead of by patient. Since all of the patients in the cohort had atypical cervical cells from previous evaluations, it was not possible to determine if there might be a difference in morphology between patients with healthy cervical tissue and those with atypical cells. Further study is needed to determine if a/LCI measurements of histologically normal appearing tissue can detect field effect of carcinogenesis or other factors such as inflammation.
Unlike the previous study, the current study did not observe a significant difference between nuclear densities in non-dysplastic vs. dysplastic biopsies but did detect a statistically significant difference between nuclear densities in low-risk vs. high-risk biopsies. According to typical cervical cytological analysis, it is reasonable to detect a more substantial density change from LSIL to HSIL as compared to that from benign tissue to LSIL because HSIL is associated not only with nuclear enlargement but also decreased cytoplasmic area [24]. This observation is consistent with the data when observing the generally lower densities seen in HSIL-classified biopsies. Because most patients in the cohort had a history of dysplasia, the lack of statistically significant difference of nuclear density between non-dysplastic and dysplastic tissue may not translate to an observed difference here. However, because the decision line between non-dysplastic and dysplastic tissue used here and in the previous a/LCI clinical study was independent of density, a difference between non-dysplastic and dysplastic tissue nuclear densities may not be necessary for tissue classification.
The results of this study show the promising capabilities of scanning a/LCI as a diagnostic modality of cervical dysplasia in the clinic, although these conclusions must be weighed against the limited cohort size and diversity. As a demonstration of a new a/LCI instrument, the study shows that the device was successfully able to retrieve nuclear information from multiple different parts of the cervix without probe relocation and provide an accurate classification of the disease state of the tissue with high sensitivity and high negative predictive value, both important characteristics for a screening tool. The success of prospectively established decision lines in the histological classification shows good progress towards the establishment of a quantitative benchmark for grading cervical biopsy sites using a/LCI. This technique is also uniquely advantageous in resource-poor settings because it does not require pathologist analysis of data and that it can be used in conservative therapies to pinpoint particular regions of the cervix that exhibit dysplasia. In addition, recent developments in low-cost optics in a/LCI could reduce the cost of future implementations of scanning a/LCI [27]. Additionally, the success of scanning a/LCI as a clinical technique will justify its development for application in future studies to detect dysplasia in other areas of the body. Future work will involve a more diverse cohort such that a clearer distinction between nuclear morphology of cervical tissues with dysplasia and those without it. The instrument itself will also need engineering improvements if it is to be employed in resource-poor settings. To enable use by operators with minimal training, use of stronger and more heat-resistant resins as well as software that is easier for a clinician to use will mitigate the type of malfunctions experienced in the study. Furthermore, higher specificity to reduce the false positive rate would be desirable in the clinical application of this instrumentation. Additional examination of patients without cervical abnormalities may have an influence the false positive rate due to factors such as the field effect of carcinogenesis. We have also explored alternative algorithms for processing Mie scattering data in the past and are currently investigating using deep learning to improve classification accuracy of our a/LCI scans.
Funding
National Science Foundation (Graduate Research Fellowship); National Cancer Institute (R01-CA16742).
Disclosures
The authors declare no conflicts of interest.
References
1. M. Arbyn, E. Weiderpass, L. Bruni, S. de Sanjosé, and M. Saraiya, “Estimates of incidence and mortality of cervical cancer in 2018: a worldwide analysis,” Lancet Global Health 8(2), e191–e203 (2020). [CrossRef]
2. M. C. M. de Kok, M. A. van der Aa, M. van Ballegooijen, S. Siesling, H. E. Karim-Kos, F. J. van Kemenade, and J. W. W. Coebergh, “Trends in cervical cancer in the netherlands until 2007: has the bottom been reached?” Int. J. Cancer 128(9), 2174–2181 (2011). [CrossRef]
3. J. Cuzick, C. Clavel, K. Petry, C. J. L. M. Meijer, H. Hoyer, S. Ratnam, A. Szarewski, P. Birembaut, S. Kulasingam, P. Sasieni, and T. Iftner, “Overview of the european and north american studies on hpv testing in primary cervical cancer screening,” Int. J. Cancer 119(5), 1095–1101 (2006). [CrossRef]
4. J. Jeronimo and M. Schiffman, “Colposcopy at a crossroads,” Am. J. Obstet. Gynecol. 195(2), 349–353 (2006). [CrossRef]
5. B. Yu, D. G. Ferris, Y. Liu, and V. K. Nagarajan, “Emerging optical techniques for detection of oral, cervical and anal cancer in low-resource settings,” Austin J. Biomed. Eng. 1(2), 1007 (2014).
6. M. Brisson, J. J. Kim, K. Canfell, M. Drolet, G. Gringas, E. A. Burger, D. Martin, K. T. Simms, E. Bérnard, M. Boily, S. Sy, C. Regan, A. Keane, M. Caruana, D. T. N. Nguyen, M. A. Smith, J. Laprise, M. Jit, M. Alary, F. Bray, E. Fidarova, F. Elsheikh, P. J. N. Bloem, N. Broutet, and R. Hutubessy, “Impact of hpv vaccination and cervical screening on cervical cancer elimination: a comparative modelling analysis in 78 low-income and lower-middle-income countries,” Lancet 395(10224), 575–590 (2020). [CrossRef]
7. D. Bzhalava, P. Guan, S. Franchesi, J. Dillner, and G. Clifford, “A systematic review of the prevalence of mucosal and cutaneous human papillomavirus types,” Virology 445(1-2), 224–231 (2013). [CrossRef]
8. J. M. M. Walboomers, M. V. Jacobs, M. M. Manos, F. X. Bosch, J. A. Kummer, K. V. Shah, P. J. F. Snijders, J. Peto, C. J. L. M. Meijer, and N. Muñoz, “Human papillomavirus is a necessary cause of invasive cervical cancer worldwide,” J. Pathol. 189(1), 12–19 (1999). [CrossRef]
9. E. M. Burd, “Human papillomavirus and cervical cancer,” Clin. Microbiol. Rev. 16(1), 1–17 (2003). [CrossRef]
10. C. W. Boone, G. J. Kelloff, and V. E. Steele, “The natural history of intraepithelial neoplasia: Relevance to the search for intermediate endpoint biomarkers,” J. Cell. Biochem. 50(S16G), 23–26 (1992). [CrossRef]
11. P. F. Escobar, L. Rojas-Espaillat, S. Tischi, C. Enerson, J. Brainard, J. Smith, N. J. Tresser, F. I. Feldchtein, L. B. Rojas, and J. L. Belinson, “Optical coherence tomography as a diagnostic aid to visual inspection and colposcopy for preinvasive and invasive cancer of the uterine cervix,” Int. J. Gynecol. Cancer 16(5), 1815–1822 (2006). [CrossRef]
12. K. Carlson, I. Pavlova, T. Collier, M. Descour, and M. F. R. Richards-Kortum, “Confocal microscopy: Imaging cervical precancerous lesions,” Gynecol. Oncol. 99(3), S84–S88 (2005). [CrossRef]
13. S. Duraipandian, W. Zheng, W. Ng, J. J. H. Low, A. Ilancheran, and Z. Huang, “Simultaneous fingerprint and high-wavenumber confocal raman spectroscopy enhances early detection of cervical precancer in vivo,” Anal. Chem. 84(14), 5913–5919 (2012). [CrossRef]
14. J. A. Freeberg, D. M. Serachitopol, N. McKinnon, R. Price, E. N. Atkinson, D. D. Cox, C. MacAulay, R. Richards-Kortum, M. Follen, and B. Pikkula, “Fluorescence and reflectance device variability throughout the progression of a phase ii clinical trial to detect and screen for cervical neoplasia using a fiber optic probe,” J. Biomed. Opt. 12(3), 034015 (2007). [CrossRef]
15. D. Ho, T. K. Drake, K. K. Smith-McCune, T. M. Darragh, L. Y. Hwang, and A. Wax, “Feasibility of clinical detection of cervical dysplasia using angle-resolved low coherence interferometry measurements of depth-resolved nuclear morphology,” Int. J. Cancer 140(6), 1447–1456 (2017). [CrossRef]
16. D. Ho, T. K. Drake, R. C. Bentley, F. A. Valea, and A. Wax, “Evaluation of hybrid algorithm for analysis of scattered light using ex vivo nuclear morphology measurements of cervical epithelium,” Biomed. Opt. Express 6(8), 2755–2765 (2015). [CrossRef]
17. Z. A. Steelman, D. Ho, K. K. Chu, and A. Wax, “Scanning system for angle-resolved low-coherence interferometry,” Opt. Lett. 42(22), 4581–4584 (2017). [CrossRef]
18. N. G. Terry, Y. Zhu, M. T. Rinehart, W. J. Brown, S. C. Gebhart, S. Bright, E. Carretta, C. G. Ziefle, M. Panjehpour, J. Galanko, R. D. Madanick, E. S. Dellon, D. Trembath, A. Bennett, J. R. Goldblum, B. F. Overholt, J. T. Woosley, N. J. Shaheen, and A. Wax, “Detection of dysplasia in barrett’s esophagus with in vivo depth-resolved nuclear morphology measurements,” Gastroenterology 140(1), 42–50 (2011). [CrossRef]
19. N. G. Terry, Y. Zhu, J. K. M. Thacker, J. Migaly, C. Guy, C. R. Mantyh, and A. Wax, “Detection of intestinal dysplasia using angle-resolved low coherence interferometry,” J. Biomed. Opt. 16(10), 106002 (2011). [CrossRef]
20. Y. Zhu, N. G. Terry, J. T. Woosley, N. J. Shaheen, and A. Wax, “Design and validation of an angle-resolved low-coherence interferometry fiber probe for in vivo clinical measurements of depth-resolved nuclear morphology,” J. Biomed. Opt. 16(1), 011003 (2011). [CrossRef]
21. H. Zhang, Z. A. Steelman, D. S. Ho, K. K. Chu, and A. Wax, “Angular range, sampling and noise considerations for inverse light scattering analysis of nuclear morphology,” J. Biophotonics 12(2), e201800258 (2019). [CrossRef]
22. W. J. Brown, J. W. Pyhtila, N. G. Terry, K. J. Chalut, T. A. D’Amico, T. A. Sporn, J. V. Obando, and A. Wax, “Review and recent development of angle-resolved low-coherence interferometry for detection of precancerous cells in human esophageal epithelium,” IEEE J. Sel. Top. Quantum Electron. 14(1), 88–97 (2008). [CrossRef]
23. D. Ho, “Development of optical technologies for comprehensive screening of cervical epithelial health,” Ph.D. thesis, Duke University (2019).
24. M. R. Henry, D. K. Russell, R. D. Luff, M. U. Prey, T. C. Wright Jr., and N. Nayar, “Epithelial cell abnormalities: Squamous,” in The Bethesda System for Reporting Cervical Cytology, R. Nayar and D. C. Wilbur, eds. (Springer International Publishing, Switzerland, 2015), 3rd ed.
25. F. E. Robles, Y. Zhu, J. Lee, S. Sharma, and A. Wax, “Detection of early colorectal cancer development in the azoxymethane rat carcinogenesis model with fourier domain low coherence interferometry,” Biomed. Opt. Express 1(2), 736–745 (2010). [CrossRef]
26. D. Damania, H. K. Roy, D. Kunte, J. A. Hurteau, H. Subramanian, L. Cherkezyan, N. Krosnjar, M. Shah, and V. Backman, “Insights into the field carcinogenesis of ovarian cancer based on the nanocytology of endocervical and endometrial epithelial cancer cells,” Int. J. Cancer 133(5), 1143–1152 (2013). [CrossRef]
27. Z. A. Steelman, D. Ho, Y. Zhao, H. Zhang, E. Jelly, G. Song, W. Y. Kendall, M. Crose, B. Cox, K. K. Chu, and A. P. Wax, “Progress in angle-resolved low-coherence interferometry for real-time detection of epithelial dysplasia (Conference Presentation),” in Biomedical Applications of Light Scattering X, vol. 11253A. Wax and V. Backman, eds., International Society for Optics and Photonics (SPIE, 2020).
