Validation of a Semiautomatic Optical Coherence Tomography Digital Image Processing Algorithm for Estimating the Tear Meniscus Height

Abstract

Purpose: To design and validate a high-sensitivity semiautomated algorithm, based on adaptive contrast image, able to identify and quantify tear meniscus height (TMH) from optical coherence tomography (OCT) images by using digital image processing (DIP) techniques.

Methods: OCT images of the lacrimal meniscus of healthy patients and with dry eye are analyzed by our algorithm, which is composed of two stages: (1) the region of interest and (2) TMH detection and measurement. The algorithm performs an adaptive contrast sequence based on morphologic operations and derivative image intensities. Trueness, repeatability, and reproducibility for TMH measurements are computed and the algorithm performance is statistically compared against the corresponding negative obtained manually by using a commercial software.

Results: The algorithm showed excellent repeatability supported by an intraclass correlation coefficient equal to 0.993, a within-subject standard deviation equal to 9.88, and a coefficient of variation equal to 2.96%, and for the reproducibility test, the results did not show a significant difference as the mean value was 244.4 ± 114.9 µm for an expert observer versus 242.4 ± 111.2 µm for the inexperienced observer (P = 0.999). The method strongly suggests the algorithm can predict measurements that are manually performed with commercial software.

Conclusions: The presented algorithm possess high potential to identify and measure TMH from OCT images in a reproducible and repeatable way with minimal dependency on user.

Translational Relevance: The presented work shows a methodology on how, by using DIP, it is possible to process OCT images to calculate TMH and aid ophthalmologists in the diagnosis of dry eye disease.

Introduction

Dry eye disease (DED) is a condition of the ocular surface represented by a loss of homeostasis, instability and hyperosmolarity of the tear film, inflammation and damage of the ocular surface, and neurosensory abnormalities.¹ According to its etiology, DED can be caused by aqueous deficiency, increased evaporation, or both. In 2017, the Dry Eye Workshop II (DEWS II) report described the presence of symptoms assessed with a validated questionnaire (Ocular Surface Disease Index [OSDI] or Dry Eye Quesstionaire-5 [DEQ-5]) plus a finding of loss of homeostasis (tear breakup time <10 seconds, hyperosmolarity or ocular surface damage). Also, for identifying aqueous deficient dry eye and evaluating its severity, the report suggests tear meniscus measurement, considering a value of 200 microns as a cutoff point for dry eye diagnosis.² 

The evaluation of tear meniscus dimension, known as meniscometry, can be done with several techniques, some of them invasive and others noninvasive. Noninvasive techniques are preferred because they reflect the natural state of the meniscus by avoiding additional fluorescein staining volumes necessary for proper detection in invasive techniques. There are several noninvasive techniques, and the most frequently used are interferometry,³ reflectometry,⁴ and optical coherence tomography,^5,6 with the latter being the most preferred since it not only allows evaluation of the tear meniscus height (TMH), tear meniscus area, tear meniscus volume, and tear meniscus radius but also is accessible as a subsystem of equipment employed in many medical offices because of its diversity of use. It is important to consider that the use of these techniques presents the disadvantage of manual measurements, resulting in subjective intra- and interobserver TMH values.⁷ 

To solve this problem, some authors have developed devices and computational algorithms that allow semiautomizing the measurement, most of them by using generic image manipulation software.⁸ 

Optical coherence tomography (OCT) images help to improve TMH measurement compared to the slit lamp, because reflections are minimized and present higher precision when an anterior eye cross-sectional image is captured. Imamura et al.⁵ studied the usability and reproducibility of the slit-lamp methods, which used a grid, versus swept-source OCT (SS-OCT). They found that the TMH values obtained by using SS-OCT images were significantly higher than those obtained by the slit-lamp method, deducing that the latter does not offer the same image resolution of the tear film as in SS-OCT. Although the slit-lamp method showed less variability than the SS-OCT method, it was also found that it can cause photophobia. Lam et al.⁶ compared TMH measurements between the Keratograph K5M and the SS-OCT SS-1000 (Oculus Optikgeräte, Wetzlar, Germany), where nonsignificant differences were found. Both kinds of equipment provided similar TMH results in healthy subjects, but as they only tested asymptomatic subjects, they concluded that it is necessary to measure TMH in patients with a DED diagnosis. 

In addition, it has been proved that using cross-sectional images of the tear meniscus captured with OCT improves differentiation between healthy subjects and patients with aqueous DED with a good sensitivity and specificity to differentiate between healthy and dry eye patients.^9,10 Canan et al.¹¹ analyzed the TMH interobserver reproducibility in Fourier domain optical coherence tomography (FD-OCT) images, by using a caliper to identify the TMH, and concluded that OCT images have a good reproducibility mainly because of due to their high resolution of the images. 

Since there are few specific algorithms and they lack repeatability and reproducibility validation in healthy and DED patients, we conducted a study that aims to design and validate a high-sensitivity algorithm, based on an adaptive contrast image, able to identify and quantify TMH from OCT images by using digital image processing techniques with a minimal user dependency. 

Methods

OCT images of the lacrimal meniscus from the files of healthy individuals and patients with dry eye from the Ocular Surface Clinic, Destellos de Luz Foundation, were included. The study protocol was approved by the institution's committee and was conducted according to the guidelines of the Declaration of Helsinki and the guidelines of Good Clinical Practice. For healthy subjects, the inclusion criteria included an Ocular Surface Disease Index (OSDI) <12 points, noninvasive tear film breakup time (NIKBUT) >10 seconds, and negative ocular surface staining in order to exclude any patient with DED. Patients with DED were included according to the criteria of the 2017 Tear Film and Ocular Surface II (TFOS II) workshop²: briefly, an OSDI >12 points and one sign of disruption to the tear film or ocular surface homeostasis, such as NIKBUT <10 seconds, ocular surface staining or tear osmolarity value ≥308 mOsmol/L in the eye with the higher osmolarity, or an interocular difference in osmolarity >8 mOsmol/L. 

Image Acquisition

Images were obtained using CIRRUS HD 5000 OCT SCAN equipment (Carl Zeiss Meditec, Inc., Dublin, CA, USA), which uses an 840-nm wavelength to produce a 9-mm image of 381 × 1247 pixels. In the absence of a specific algorithm for the evaluation of TMH, the algorithm for corneal evaluation (which is included by the CIRRUS system) was chosen combined with a 60-diopter lens and the measurement of 1 high-definition (HD) line, reaching a high contrast regardless of the large size of the image. To acquire the images of the tear meniscus, the technique described by Bitton et al.¹² was employed. Before the measurement started, the ambient light was reduced, and the patient was asked to blink normally to avoid a reflex tearing. The patient was placed with their gaze straight ahead, and the vertical measurement marker was placed on the VI meridian of the cornea, including the peripheral cornea and the lid margin in the measurement (see Fig. 1). The patients were asked to blink again, and the measurement was taken 3 seconds later, guided by an audible alarm with a stopwatch started when the patient opened their eyes, and a beep alarm was set to 3 seconds to have a stable tear film and meniscus. Inferior TMH was defined as the distance between the corneal–meniscus junction and lower eyelid–meniscus junction. The meniscus height was then determined manually by the examiner by using the software included in the CIRRUS HD-5000 equipment. 

Figure 1.

 

Scan alignment.

The same measurement was performed by our designed semiautomatized algorithm, which takes advantage of the high-resolution images captured by the mentioned OCT SCAN equipment. The digital image processing was made by using the programming software MATLAB, R2020b (The MathWorks, Natick, MA, USA), based on two stages: (1) region of interest (ROI) selection and (2) TMH detection and measurement. 

ROI Selection

Because of the large image dimensions, a method for setting the ROI over the images to be analyzed was designed (see Fig. 2). 

Figure 2.

 

ROI selection algorithm.

Initially, the user selects the lowest point of the tear meniscus to later display an automatic crop of the original image. At this step, the user must decide whether to readjust the image crop or to continue with the TMH measurement algorithm, based on the appearance of the fitted image. Figure 3 shows several automatic image crop dimensions, with the best options (Figs. 3A–C) those that approximate an x/y aspect ratio of 1.2, where “x” represents the horizontal axis and “y” the vertical axis. If the image does not approximate this ratio or the meniscus extension does not appear completely within the proposed image, a manual option is enabled to reach an appropriate result. At this point, a list containing the recommended characteristics for the image crop is displayed as a dialog box, giving a user's guide for the ROI selection process. 

Figure 3.

 

Automatic image crop examples. (A–C) Correct meniscus appearance. (D) Incomplete meniscus. (E, F) Small meniscus appearance.

Once the final ROI is determined, the resultant image is employed as the input of a second stage: the TMH detection and measurement. 

TMH Detection and Measurement

For the TMH measurement algorithm, we used the resultant image of the ROI selection algorithm as input (Fig. 4A). The processing is divided into four steps that are described below and whose parameters were obtained after an iterative process: (1) binarizing and edge detection, (2) removal of small objects and morphologic operations, (3) count-controlled loop with conditions, and (4) unit conversion. 

Figure 4.

 

TMH detection and measurement algorithm.

Binarizing and Edge Detection

First, the input image is binarized by classifying the intensity level of each pixel in five threshold levels using Otsu's method¹³ (Fig. 4B). The first class is used as the thresholding value, because of the low image contrast. Then, a Sobel digital filter is used horizontally as well as vertically to detect edges in the binarized image. 

Removal of Small Objects and Morphologic Operations

In this stage, noise represented by binary large objects is eliminated by removing those objects detected with a size of fewer than 20 pixels (Fig. 4C). Next, a closing morphologic tool is used to unite the resultant triangular area that contains the tear meniscus region (Fig. 4D). 

Count-Controlled Loop With Conditions

Knowing that TMH is overestimated because of the tilts formed in the corneal and/or eyelid region, it is necessary to focus on the central portion. The image is analyzed by dividing it into two regions, and then a horizontal sweep is made from the center to the sides in both sections (Fig. 4E). The first nonzero pixel of each column in the image is detected by scanning from top to bottom of representative (white) pixels of the image (Fig. 4F). To select and count the pixels that belong to the TMH, the spatial derivative of the image is made, and the pixel sequence is analyzed along the mentioned direction by a local window (Fig. 4G). 

Unit Conversion

Once the number of resultant pixels is registered in the count-controlled loop, a conversion to microns is made considering the dimensions of the original image according to the instrument’s optical system, which are established in the technical specifications of the capture equipment as 7.21 µm per pixel (Fig. 4H). The previous value corresponds to a fixed 9-mm horizontal field of view, divided by a horizontal image size of 1247 pixels (see image acquisition section); in this way, because the optical setup is fixed in the OCT equipment, whenever a focused image is captured, the unit conversion is valid. 

Trueness, Repeatability, and Reproducibility

The algorithm’s trueness assessment was made by comparing the results of two methods: (1) the measurements made with our TMH algorithm evaluated by an expert observer (E1) versus (2) the measurement made by another expert observer employing the manual OCT CIRRUS software (E2). 

To evaluate the repeatability of the algorithm, the observer (E1) made five measurements with a difference time of a week between measurements. We applied a randomization in the sequence of images each week before (E1) executing the test. 

Interobserver reproducibility was evaluated by comparing the measurements between the expert (E1) and a nonexpert observer (E3) who developed the TMH algorithm under the same randomization procedure. 

Sample Size

Using 15 images, a pilot test was conducted in order to get the standard deviation of the TMH measurement. Considering the results, the full study required a minimum sample size of 57 images to achieve a power of 95% and a level of significance of 5% (two-sided), for detecting a mean of the differences of 2.5 microns between pairs, assuming the standard deviation of the differences to be 5.06 microns. For the sample size required for precision studies, based on the number of repeated measures and the confidence in the estimate reported by McAlinden et al.,¹⁴ with a set of 60 images, a power of 95% and a level of significance of 10% are achieved. 

Statistical Analysis

Statistical analysis was made using SPSS (IBM SPSS Statistics Version 25.0, SPSS, Inc., Chicago, IL, USA). Comparisons between the CIRRUS-OCT software and our algorithm were conducted using the following: 

•  
  Comparison of means. To compare the means of all measurements, a Wilcoxon matched-pairs signed rank test was used.
•  
  A Bland–Altman plot was also constructed to compare the results from the two methods using Prism 9 for macOS (GraphPad Software, San Diego, CA, USA). Differences between the CIRRUS-OCT software and our algorithm are plotted against the average of them. The limits of concordance were calculated as the average difference of the measurements with each method ±1.96 the standard deviation of the differences. By definition, a 2.00 standard deviation is the concordance range between techniques, and the lower the value, the higher the concordance.

Repeatability was assessed using three indices: 

• Intraclass correlation coefficient (ICC). ICCs were calculated for each method by dividing the between-subjects variance by the sum of the between- and within-subjects variance (ICC = (σ2b/(σ2b + σ2w))). The between-subjects variance indicates differences observed across individuals, while the within-subjects variance reflects differences across measures in the same individual (or “error”). Thus, the ICC estimates the proportion of the total variance that is attributable to differences between subjects (“true” variance), being close to 1 when no variance between measures in the same individual exists, reflecting high repeatability.
• Within-subject standard deviation (Sw). Standard deviation was calculated for the three measures of each individual and then pooled across individuals. The pooled within-subjects standard in this study represents the average expected deviation when measuring any single subject on three occasions.
• Coefficient of variation (CV). The CV was calculated by dividing the pooled within-subjects standard deviation by the pooled within-subjects mean. The CV is a standardized measure of variability that allows comparisons between data sets with different means. A smaller CV reflects better repeatability. CV was expressed as a percentage to ease comparisons.

For the calculation of the standard deviation, three serial measurements were made. Between each measurement, the subject detached from the equipment blinked naturally for 30 seconds, and the process previously described for image acquisition was repeated. 

For the comparison between measurement techniques, we used the first measurement of both the observer and the algorithm to evaluate the robustness of the second. 

An analysis was performed to identify outliners with the ROUT method of GraphPad Prism, and an outliner was identified, with simple linear regression analysis and correlation performed after eliminating the result. 

Results

The study included 266 images of 79 patients, with a mean age of 54 ± 4 years, 70% of them (56) women. The right eye was evaluated in 62.5% of the cases (50 images). Images from 60 patients were used because 19 patients (24.05%) presented an image that could not be measured correctly, mainly because of a specular reflection (overexposed images) that formed a vertical line in the center of the ROI and caused the algorithm to stop because it did not meet the loop conditions, causing it to take mistake the specular reflection as the cornea because of its vertical shape. This can be seen in Figure 5. 

Figure 5.

 

Vertical line preventing correct TMH measurement that occurred due to an orthogonal alignment between the observing axis of the equipment and the corneal surface of the patient.

Algorithm repeatability was excellent, with an ICC of 0.993 (95% confidence interval [CI], 0.989–0.995). The resulting Sw and CV were 9.88 and 2.96%, respectively. The mean values for the algorithm measurement in the repeatability test were 248.4 ± 115.1 µm, 244.2 ± 112.1 µm, 247.4 ± 115.8 µm, 246.6 ± 151.4 µm, and 246.2 ± 118.1 µm. The interobserver reproducibility did not show a significant difference, as the mean value was 244.4 ± 114.9 µm for the expert observer (E1) vs. 242.4 ± 111.2 µm for the inexperienced observer (E3) (P = 0.999). 

Bland–Altman plots showed an adequate agreement between both observers (E1 and E3), since bias is very close to zero and data are dispersed between ±1.96 SD, showing no pattern that indicates bias for the algorithm-experienced observer (Fig. 6A) and the observer with no experience with the algorithm (Fig. 6B). 

Figure 6.

 

Bland–Altman plots. (A) Expert observer (E1). (B) Inexperienced observer (E3).

A simple linear regression method was also used to identify if the proposed algorithm predicts the manual measurement of (E2) with the CIRRUS software; the regression results (Fig. 7) suggest that there is a positive and statistically significant correlation (P < 0.001). 

Figure 7.

 

Linear regression showing a strong correlation between the proposed algorithm and the manual measurement with CIRRUS software.

Discussion

The algorithm programmed in this work can identify and quantify TMH from DIP, using images captured with the CIRRUS HD 5000 OCT equipment (Carl Zeiss Meditec, Inc.), which uses an-840 nm wavelength to produce a 9-mm image of 381 × 1247 pixels. This is promising as a proof of principle since it accomplishes the requirements to be accepted in the clinical area, such as minimal dependence on the user compared to manual measurements, reproducibility, repeatability and a 95% confidence level.¹⁵ 

The results of this algorithm are comparable to those obtained by equipment currently used in terms of precision. Baek et al.¹⁶ evaluated the agreement with the FD-OCT measurements and the Keratograph 5M, and they found that the TMH measured with both methods was well correlated. The intraobserver assessment with OCT images measured with a built-in caliper showed an ICC of 0.992 and a CV of 6.9%, and with the Keratograph 5M, an ICC of 0.987 and a CV of 6.43% were reported. In the present work, the corresponding ICC and CV resulted in 0.993 and 2.96%, respectively, improving the precision and variation of these indicators, as seen above. The latter is attributed to the semiautomation of the algorithm. Because of the dynamic nature of the tear film, several images must be analyzed to get a reliable TMH measurement with minimum user intervention along the process¹⁷; however, it is necessary to remember that the present algorithms do not aim to evaluate TMH dynamics but the correct TMH length measurement. In the present algorithm, the user takes part in only 2 of the 13 steps, giving an acceptable reliability according to the previous indicators. 

There are studies that used interferometric equipment and compared their measurements against those obtained by swept-source and ultra-high-resolution OCT image processing in a semiautomatic and automatic procedure, respectively, but a fully validated and automated method that uses OCT images to calculate TMH as a parameter to aid in the diagnosis of DED is not currently reported, so this work represents a necessary step in the verification of the implied clinical technologies for their quantification and diagnosis.^18,19 Another advantage of the present algorithm is that it is robust enough to work with images as small as 110 × 80 pixels without compromising precision on the results. This is because it achieves and dynamically enhances a contrast of the tear meniscus in the image for processing, largely, thanks to the adaptive filtering used. 

The use of the algorithm must consider the following limitations. The images used were not classified by severity of DED and correspond to unique equipment (CIRRUS HD 5000), so the parameters used were found based on an iterative process for the images delivered by this specific model. However, because of its high sensitivity, the technique used has the benefit of being able to be adapted easily to different characteristics of other equipment in terms of spatial resolution, bit depth, and image size. Regarding the eight images that were eliminated because of causes inherent to the capture equipment, a warning should be added to the user in a future version to ask for a new image of the patient, since the specular reflection effect was presented in 1 of every 10 images to be evaluated approximately. Regarding the other 11 images that were removed from the sample because of the disparity between the x/y dimension ratio in the step of cropping, it is recommended that a future version of the algorithm gives a warning to the user if the manual cropping performed by the user has a nonideal dimension that avoids an adequate measurement of the TMH. 

Conclusion

The presented algorithm has high potential to identify and measure TMH in a reproducible and repeatable way with minimal user dependency. For the type of OCT images delivered by the used equipment (CIRRUS HD 5000), the proposed algorithm, which is based on adaptive techniques to improve the contrast in the ROI of the studied OCT images, can be performed successfully during computer processing, so it can be employed by ophthalmologists in the clinical area to calculate TMH as a parameter to aid in the diagnosis of DED. 

Acknowledgments

Supported by the Research Department of the University of Monterrey. 

Disclosure: A. Cardenas-Morales, None; M.F. Tamez-Olvera, None; M.P. Cervantes-Rios, None; M. Garza-Leon, None; M. Tomasi, None; C.G. Tavera-Ruiz, None