Downloaded from http://journals.lww.com/corneajrnl by BhDMf5ePHKav1zEoum1tQfN4a+kJLhEZgbsIHo4XMi0hCywCX1AWnYQp/IlQrHD3i3D0OdRyi7TvSFl4Cf3VC4/OAVpDDa8K2+Ya6H515kE= on 03/15/2021 Downloadedfrom http://journals.lww.com/corneajrnlby BhDMf5ePHKav1zEoum1tQfN4a+kJLhEZgbsIHo4XMi0hCywCX1AWnYQp/IlQrHD3i3D0OdRyi7TvSFl4Cf3VC4/OAVpDDa8K2+Ya6H515kE=on 03/15/2021
Novel Application Software for the Semi-Automated
Analysis of Infrared Meibography Images
Danish Shehzad, PhD,* Sona Gorcuyeva, MD,
† Tamer Dag, PhD,* and Banu Bozkurt, MD, MSc†
Purpose: To develop semi-automated application software that quickly analyzes infrared meibography images taken with the CSO Sirius Topographer (CSO, Italy) and to compare them to the manual analysis system on the device (Phoenix software platform).
Methods: A total of 52 meibography images verified as high quality were used and analyzed through manual and semi-automated meibomian gland (MG) detector software in this study. For the manual method, an experienced researcher circumscribed the MGs by putting dots around grape-like clusters in a predetermined rectangular area, and Phoenix software measured the MG loss area by percentage, which took around 10 to 15 minutes. MG loss was
graded from 1 (,25%) to 4 (severe .75%). For the semi-automated
method, 2 blind physicians (I and II) determined the area to be masked by putting 5 to 6 dots on the raw images and measured the MG loss area using the newly developed semi-automated MG detector application software in less than 1 minute. Semi-automated measurements were repeated 3 times on different days, and the results were evaluated using paired-sample t test, Bland–Altman, and kappa k analysis.
Results: The mean MG loss area was 37.24% with the manual
analysis and 40.09%, 37.89%, and 40.08% in thefirst, second, and
third runs with the semi-automated analysis (P , 0.05). Manual
analysis scores showed a remarkable correlation with the semi-automated analysis performed by 2 operators (r = 0.950 and r =
0.959, respectively) (P , 0.001). According to Bland–Altman
analysis, the 95% limits of agreement between manual analysis
and semi-automated analysis by operator I were between210.69%
and 5% [concordance correlation coefficient (CCC) = 0.912] and
between29.97% and 4.3% (CCC = 0.923) for operator II. The limit
of interoperator agreement in semi-automated analysis was between 24.89% and 4.92% (CCC = 0.973). There was good to very good agreement in grading between manual and semi-automated analysis results (k 0.76–0.84) and very good interoperator agreement with
semi-automated software (k 0.91) (P, 0.001).
Conclusions: For the manual analysis of meibography images, around one hundred dots have to be put around grape-like clusters to determine the MGs, which makes the process too long and prone to errors. The newly developed semi-automated software is a highly reproducible, practical, and faster method to analyze infrared meibog-raphy images with excellent correlation with the manual analysis. Key Words: infrared meibography, meibomian glands, automatic detection, correlation, Kappa statistic
(Cornea 2019;38:1456–1464)
T
he meibomian glands (MGs) are modified sebaceous glands located in the tarsal plates of the eyelids and consist of excretory acini connected with long central ducts via short ductules.1 They secrete meibum, which functions as a barrier to prevent excess evaporation of the water component of the tearfluid and therefore is essential for the stability of the tear film. MG dysfunction (MGD) is defined as a chronic, diffuse abnormality of the MGs characterized by obstruction of the excretory ducts and orifices and/or qualitative/quantitative changes in the glandular secretion.2MGD may be caused as a result of aging, contact lens use, dermatologic disorders, Stevens–Johnson syndrome, or chemical burn of the ocular surface.3,4 It leads to evaporative type dry eye and ocular surface inflammation.Determination of the changes in the acini of the MGs is important for the diagnosis and management of MGD in the clinical setting. Slit-lamp biomicroscopic examination is a non-invasive examination of the eyelids for the evidence of irregular lid margins, telangiectasia and hyperemia at the orifices, meibum orifice plugging, decreased expressibility, and provocation of meibum expression through application of digital pressure to the eyelid.1–3MGD is associated with decreased tear break-up time, decreased lipid layer thickness, increased tear osmolarity, and increased Ocular Surface Disease Index scores.
The morphology of the MGs can be evaluated by in vivo diagnostic devices such as meibography using infrared (IR) light,5–13optical coherence tomography14, and confocal micros-copy.15 Meibography is a specialized imaging technology developed exclusively for visualization of the morphology of the MGs in vivo.9 It enables detection of MG dropout, shortening, dilation, and distortion. There are numerous instru-ments offering noncontact meibography on the market, includ-ing the Topcon BG-4M for slit lamp, the Meibom Pen (Japan Focus Corporation, Tokyo, Japan), and topography devices such as the Eye Top Topographer, CSO Sirius Topographer (Flor-ence, Italy) and Oculus Keratograph 5M (Oculus, Wetzlar,
Received for publication December 13, 2018; revision received June 24, 2019; accepted June 29, 2019. Published online ahead of print September 2, 2019.
From the *Computer Engineering Department, Kadir Has University, Istanbul, Turkey; and†Department of Ophthalmology, Selcuk University, Konya, Turkey.
The authors have no funding or conflicts of interest to disclose.
Correspondence: Danish Shehzad, PhD, Computer Engineering Department, Kadir Has University, Istanbul 34083, Turkey (e-mail: danish.shehzad@ khas.edu.tr).
with a Scheimpflug camera. It provides data for corneal thickness, elevation, curvature, and power of both corneal surfaces over a diameter of 12 mm, anterior chamber depth, lens thickness, pupillography, and IR meibography. Sirius uses the Phoenix software platform and allows the patient’s data to be saved for future review and analysis. The aim of this work is to develop novel software to determine the percentage of MG loss in IR meibography images taken with the Sirius Topography device and compare its results with the manual analysis system on the device. In this study, we also determined intraoperator and interoperator reproducibility by reanalyzing the IR meibography images in 3 different days by 2 examiners.
METHODS
The study was conducted after the approval of the Selcuk University Faculty of Medicine Ethical Committee (2017/245). Two hundred images taken with the CSO Sirius Topographer were examined, and 52 images (26 of upper eyelids and 26 of lower eyelids) verified in high quality with good focus and proper eyelid eversions were entered into the study.
Manual Analysis
The same researcher (S.G.) marked the borders of the tarsus trying to cover at least 90% of the upper and lower eyelids and circumscribed the MGs by putting around one hundred dots around grape-like clusters. The green zone was the area of the MGs, and the red zone was the MG loss area. Afterfinishing, Phoenix software gave the measurements of dropout both by percentage and by a scale within the area, which was highlighted by the users’ free-hand tool; grade 1: ,25%, grade 2: 25% to 50%, grade 3: 50% to 75%, and grade 4:.75%.8,9 Time for analysis and percentage of loss area were noted for each image. The raw images with predetermined areas used in the manual analysis were also sent to the researchers of the Computer Engineering Depart-ment (T.D. and D.S.), who were masked to the results of manual analysis.
Newly Developed Application Software—
Semi-Automated MG Detector
The semi-automated MG detector was developed and tested by the authors of this work. For the development of the semi-automated software, the matrix laboratory (MATLAB) platform is used. MATLAB (https://www.mathworks.com) is considered as a multiparadigm numerical computing environ-ment.16 It allows a user to implement algorithms and plot functions and data. The toolboxes that come along with MATLAB may be used in a broad-spectrum of areas such as statistics and image processing.
area, image processing techniques were used in MATLAB. The procedure showing how to use the newly developed semi-automated software is described in the following steps:
Loading Image
In the first step, by clicking on the “Load Image” button, the user searches for the location of the image and loads it to the application software to start the processing (Fig. 1A).
Region of Interest Selection
The region of interest (ROI) selection is activated by clicking on the “Select ROI” button (Fig. 1B). The ROI is selected using the image-free hand tool, and the area is kept polymorphic to achieve accurate results. The users can determine the ROI by placing 5 to 6 dots on the image through mouse clicks. This step takes no more than a few seconds.
Median Filtering
After selection of the ROI, the medianfilter, which is a nonlinear digital filtering technique, is applied on the imagefile to remove noise from the image by clicking on the“Median Filter” button (Fig. 1C). Such noise reduction is a typical preprocessing step to improve the results of later processing. It filters the image matrix in 2 dimen-sions. An output pixel encompasses the median value in a 3 · 3 neighborhood around the corresponding pixel of the input image.
Adaptive Thresholding
Adaptive thresholding uses localfirst-order statistics, applied to thefiltered image (Fig. 1D). It takes into account spatial variations in illumination, and it converts an intensity image into a binary image. Unlikefixed threshold, the threshold value at each pixel location depends on the neighboring pixel intensities. The threshold value can be adjusted using the slider according to the requirements to achieve accurate results. MGs are marked as white zones and dropout areas as black zones.
After following the above steps, the software automat-ically calculates and displays the percentage of the MGs and dropout areas in the selected ROI. Based on the results, the grade of MGD is determined. The time spent between the loading of the image and obtaining the results takes less than 1 minute.
Interoperator reproducibility of the application software was evaluated by comparing the measurements taken by 2 physicians (I and II). Intraoperator reproducibility was evaluated by repeating the measurements on 2 different days by the same physician (I).
Statistical Analysis
Statistical analysis was performed with the commer-cial software SPSS for Mac, version 23 (SPSS Sciences, Chicago, IL). All the data are expressed as mean 6 SD. The normal distribution of variables was verified by the Shapiro–Wilk test.
Manual and automated analysis by the new software was compared by the paired-sample t test, and the agree-ments between the measureagree-ments were evaluated using the Bland–Altman and kappa k analysis. The Bland–Altman analysis is considered as a graphical tool to compare 2 measurement techniques. A Bland–Altman plot shows the differences between the 2 techniques versus the averages of the 2 techniques. The horizontal lines show the mean dif-ference (MD) and the limits of agreement, which are defined as MD 6 1.96· the SD of the differences. The concordance correlation coefficient (CCC) shows the agreement of 2 techniques in statistics, and a value of 1 denotes perfect agreement. Correlations between variables were undertaken using the Pearson correlation analysis, depending on the distribution of the variables. P , 0.05 was considered statistically significant.
RESULTS
Fifty-two images of 39 patients (26 female, 13 male, and age range 19–79 years) were studied. Twenty-one subjects (11 female and 10 male) had MGD, and 18 subjects (15 female and 3 male) were healthy. The mean ages of patients with MGD and healthy subjects were 47.9 6 17.05 years and 47.86 17.7 years, respectively, with no statistically significant differences (P = 0.97). There were 26 lower eyelid
meibography images (15 healthy and 11 MGD) and 26 upper eyelid images (6 healthy and 20 MGD).
The software calculated the ratio of the total MG area relative to the total analysis area in all images. The mean MG loss area was 37.24%6 12.34% in manual analysis, 40.09% 6 10.59% in the first analysis of operator I, 37.89% 6 12.17% in the second day analysis of operator I, and 40.08% 6 10.82% in the third day analysis of operator II, and using the paired-sample t test, the mean MG loss area in manual analysis was statistically significantly lower than automated analysis measurements (P , 0.05). The mean time for processing of detection of the MG loss area in manual analysis was 15 6 3.4 minutes, whereas it took less than 1 minute on average by applying all the steps described above with automated analysis.
Figure 2 shows some meibography images analyzed both with manual system and semi-automated software. Two of these images were taken from the upper eyelids and 2 from the lower eyelids. The MG loss area measurements of both systems are very similar; however, the semi-automated software percentage results are usually higher than the manual results. The reason of this difference is due to the fact that the observer is more likely to skip some minor areas between the glands when putting dots around, whereas the semi-automated software is able to catch them correctly (Fig. 3). A limitation of the new software is that it catches the light reflections and scar tissues on the images and classifies these areas as the MGs (Fig. 4). Although the differences are most of the time negligible, the new version of the software will try to overcome those problematic regions by adding a new component in the software that will help to omit regions such as light reflections and scar tissues.
FIGURE 1. Stepwise semi-automated meibomian gland detection process. A, Loading image. B, Select region of interest (ROI). C, Apply median/mean filtering. D, Adapative thresholding.
Figure 5 shows the correlation between manual detec-tion of the MGs and automated detecdetec-tion of the MGs by the 2 examiners, respectively. The Pearson correlation test yielded a correlation coefficient of r = 0.950 (P , 0.001) between manual and semi-automated analysis by operator I (Fig. 5A). Similar results have also been obtained for operator II with a Pearson correlation coefficient of r = 0.959 (P , 0.001) (Fig. 5B).
Figure 5C shows the correlation between the semi-automated analysis results of operators I and II (r = 0.973, P , 0.001), and Figure 5D shows the correlation between the semi-automated analysis results obtained by the same oper-ator (operoper-ator I) on 2 different days (r = 0.951, P, 0.001). We performed the Bland–Altman analysis to determine the consistency between the manual and semi-automated test
results (Fig. 6). The 95% limits of agreement between manual analysis and semi-automated analysis by operator I span from 210.69% to 5% with an MD of 22.85 6 4 (CCC = 0.912) (Fig. 6A), and the 95% limits of agreement between manual analysis and semi-automated analysis by operator II spans from29.97% to 4.3% with an MD of 22.83 6 3.64 (CCC = 0.923) (Fig. 6b). The 95% limits of agreement between operators I and II (interoperator) were found to span from 24.89% to 4.92% with an MD of 0.01% 6 2.5 (CCC 0.973) (Fig. 6C) and 25.39% to 9.78% with an MD of 2.19% 6 3.87 (CCC = 0.925) for semi-automated analysis test results obtained by operator I (intraoperator) on 2 consecutive days (Fig. 6D).
The Bland–Altman analysis illustrates that the discrepancies between the manual analysis versus
FIGURE 2. Meibography images analyzed with the manual system and the semi-automated meibomian gland detector software. A, Original Image. B, Manual (47.9%). C, Semi-Automated (49.48%). D, Original Image. E, Manual (50.9%). F, Semi-Automated (53.36%). G, Original Image. H, Manual (33.2%). I, Automated (37.32%). J, Original Image. K, Manual (17.2%). L, Semi-automated (23.67%).
semi-automated analysis and semi-automated analysis for different operators are not significant. The semi-automated software tends to give slightly higher MG loss area percentages compared with manual analysis. For example, between manual analysis and semi-automated analysis (Fig. 6A), MD was found to be 22.85%, and the 95% limits of agreement spanned from 210.69% to 5%. The reason for semi-automated software producing higher MG loss areas is due to the fact that during manual analysis, it is very likely to skip the minor areas between the glands, whereas the semi-automated software can capture and identify these areas.
In manual analysis, 45% of the measurements were grade 1, 45% were grade 2, and 10% were grade 3 (Table 1). On the first day analysis by operator I, 25% were grade 1, 65% were grade 2, and 10% were grade 3, and on the second day analysis of operator I, 35% were grade 1, 55% were grade 2, and 10% were grade 10. On the third day analysis by operator II, 20% were grade 1, 65% were grade 2, and 15% were grade 3.
The agreements between manual and automated anal-ysis were given in Table 2. The agreement between the semi-automated analysis and the manual analysis revealed good to very good agreement (k 0.84 for I and k 0.76 for II, P , 0.001), the interobserver agreement for I and II using the new semi-automated software resulted in very good agreement (k 0.91, P, 0.001), and the intraobserver agreement for the semi-automated software by the same observer I on different days resulted in very good agreement (k 0.88, P, 0.001).
DISCUSSION
In recent years, IR imaging of the MGs and scoring systems have been widely used for the diagnosis of MGD in various disorders. With the use of meibography, the structure of the MGs, including the ducts and acini, can be observed and assessed easily. In the upper lids, the MGs appear to be thinner and longer compared with the lower eyelids, which are broader and shorter. Using meibography, we can observe pathological changes related to MGD, including gland distortion, shortening, and dropout. Gland distortion is tortuousity of glands and/or discordant pattering of glands. Shortening is gland not extending from the eyelid margin to the opposite edge of the tarsal plate, and dropout is a zone of MG loss.
Meiboscore is a semiquantitative evaluation putting MG loss in categories with various cutoff values according to different classifications. Arita et al5 proposed a practical scoring system for MG loss from 0 to 3 based on the extent of eyelid involvement: grade 0, no significant eyelid involve-ment; grade 1,,33% involved; grade 2, 33%–66% involved; and grade 3,.66% involved. The MG dropout area might also be classified into 5 groups: grade 0: no loss at all, grade 1:,25% involved, grade 2: 25%-50%, grade 3: 51% to 75%, and grade 4: .75%.8,9 The main problem with the grading systems can be explained as follows: According to the Arita scoring system, eyes with almost similar MG loss percentages such as 32% and 34% can be classified into different grades (grade 1 and 2, respectively), while a large percentage difference such as 35% and 65% will classify eyes in grade 2. It is also not proper to use these grading systems in follow-up
FIGURE 3. Inclusion of minor ef-fected areas as MG by semi-auto-mated meibomian gland detector as compared to manual detection method. A, Manually calculated image result. B, Semi-automated analysis result.
FIGURE 4. Semi-automated meibo-mian gland detector detects light reflections as MG areas. A, Original Image. B, Semi-automated detector image.
examinations because although there is a remarkable improvement in MG morphology after treatment, the eye might still be classified in the same grade (eg, the MG loss area might decrease from 60% to 35%, but still remains in grade II according to the Arita classification5), which means that there is no improvement. Therefore, instead of grading systems, the physicians better use the MG loss area percent-age for comparison and follow-up, which is more quantitative and can show small differences.
In this study, we used the CSO Sirius Topographer, an anterior segment analysis system, which combines placido disk technology with a Scheimpflug camera, that provides corneal tomographic parameters, but also allows visual documentation and digital analysis of MGD using the IR system. However, using Phoenix software, it takes at least 10 to 15 minutes to put hundreds of dots along the border of the MGs and measure the loss area as percentage, especially when the ducts are not linearly located. Sometimes, these dots combine altering all the drawing, leading to false measurements. In such a case, the operator needs to restart the whole procedure. In addition, in tiny places or around irregular ducts, it is sometimes impossible to put dots in correct places, leading to larger or smaller MG areas. In the manual analysis, the shallow areas between the glands are most of the time skipped (very difficult to put dots around
the glands, which are so close together or very curved). Therefore, we developed semi-automated MG detector application software that automatically analyzes IR mei-bography images in a fast, objective, and quantitative manner. The process of semi-automated software takes less than 1 minute on average by applying all the steps described above in the Materials and Methods section. When we compared the results of semi-automated analysis with the manual analysis using Phoenix software, we noticed that the calculated MG loss area with the applica-tion software was a few percent higher compared with the manual system.
With semi-automated software, the areas between the ducts are all included as black (dropout) areas, which are sometimes hard to include using manual clicks with Phoenix software. Semi-automated analysis test results showed an excellent correlation with the manual analysis scores (r = 0.95) (P, 0.001). According to the Bland–Altman analysis, the 95% limits of agreement between manual analysis and semi-automated analysis by operator I were between 210.69% and 5.00% (CCC = 0.912) and between 29.97% and 4.30% (CCC = 0.923) for operator II. The interoperator and intraoperator agreement in semi-automated analysis was also excellent. There was good to very good agreement in grading between manual and semi-automated analysis results
FIGURE 5. The correlation between manual and semi-automated meibomian gland detector test results. A, Operator I test results Manual vs. Semi-automated Detector (operator I). B, Manual vs. semi-automated meibomian gland detector operator II test results. C, Semi-automated meibomian gland detector results of operator I and II. D, Semi-automated meibomian gland detector results obtained by operator I on two different days.
(k 0.76–0.84) and very good interoperator agreement with semi-automated software (k 0.91) (P, 0.001). With the new software, there is no grade 0 because it takes into account the interglandular space in contrast to the Arita scoring system, which is practical but a rough evaluation, putting clinically different eyes into the same groups, and is not a good way to follow the patients. In this study, we also took interglandular spaces as the loss area with the manual system putting dots to the lateral walls of the glands. Our aim was to evaluate whether the new software would determine the MGs and free spaces close to the manual analysis and showed that it shows an excellent correlation and high reproducibility. In our following study, we will compare the semi-automated measurements of the eyes with MGD to the healthy eyes
and make an ROC analysis tofind out a cutoff value for the discrimination between diseased and healthy eyes.
In the literature, there are also some studies which can be classified as semi-automated and used the image editing software ImageJ (National Institutes of Health; http://imagej. nih.gov/ij) in analyzing the IR meibography images.19–22In the study of Pult and Riede-Pult,19–21the MGs were captured by the IR video camera, and after determining the tarsal area to be measured, the border of the MGs was marked manually by mouse clicks with the help of ImageJ software without entering the spaces between the ducts. Then, the area of the MG dropout and its ratio to the total area were calculated. The limitation of the analysis system used in the study by Pult is that it still needs an operator to determine the border of the
FIGURE 6. Limit of agreement plot showing the consistency between manual and semi-automated meibomian gland detector test results. A, Manual vs Semi-automated (operator I). B, Manual vs. Semi-automated (operator II). C, Inter-operator Repeat-ability. D, Intra-operator RepeatRepeat-ability.
TABLE 1. Distribution of MGD Grades According to Manual and Semi-automated Detector Test Results
Analysis/Grades Manual Semi-automated Day 1 (Operator I) Semi-automated Day 2 (Operator I) Semi-automated Day 3 (Operator II) I 17.31% 9.62% 13.46% 7.69% II 65.38% 73.08% 67.31% 73.08% III 17.31% 17.31% 19.23% 19.23% IV 0% 0% 0% 0%
MGs and skips loss areas between the glands. Ban et al22 conducted a detailed morphometric assessment of the MGs in 37 subjects using noncontact IR meibography and investi-gated the relationship between MG morphology and dry eye parameters and MG morphology associated with aging and sex. They used ImageJ software in analyzing the meibog-raphy images. They selected a central area of 10 mm in length and 13 mm in width and measured the mean value of the length of 3 to 5 selected central MG ducts. White images of the lesions were converted into black images, and the percent area of the MGs within the rectangle was calculated.
Koh et al23applied several algorithms and filters such as Gaussian smoothing, clustering using FIFO, dilation, filling, and pruning to obtain the gland and intergland lines of the image. Their study focused on the images of upper eyelids. Features based on gland lengths and widths were used to accurately distinguish healthy and unhealthy meibo-graphs. However, gland area computation and grading were not performed. Arita et al7 developed software for meibog-raphy to objectively evaluate the MG areas. The noninvasive meibography system they used comprised a slit lamp (SL-D7, Topcon, Tokyo, Japan) equipped with a BG-4M and a 0.5-inch CCD camera (XC-EI-50, Sony, Tokyo, Japan), an external monitor, and a recording device. Images were obtained with this system using an IR light source. By applying Wallis and Gaussian filters, the low-contrast areas of the raw image are emphasized, and the noise is reduced. After detection of the measurement area, the MGs are detected by high-passfiltering using the fast Fourier transform and g correction process. They also examined the intraoperator repeatability of image analysis with the software on 14 upper and 14 lower eyelids of 10 normal subjects and 22 upper and 22 lower eyelids of 22 patients with MGD. The 95% confidence interval of the difference in the MG area relative to the total analysis area in the normal upper eyelids, normal lower eyelids, upper eyelids of patients with MGD, and lower eyelids of patients with MGD was20.054 to 0.382, 20.397 to 20.027, 20.045 to 0.242, and20.241 to 0.062, respectively. However, they did not compare their results with the manual methods.
There are certain limitations in our newly developed application software like the manual system. Inadvertent lid distortion, eyelid folds, focus problems, or an altered vertical
the system is that it sees light reflections and wet areas as white areas and calculates within the MG area. Therefore, the eyes should be dried, and in case of any light reflection, the process should be repeated to minimize this effect. The third problem we faced in some of the images is that the scar tissues, which appeared as white areas in rough meibography images, and accepted as the MGs by the system, giving a lesser dropout percentage compared with manual analysis.
In future studies, we will try to overcome these problems using more samples andfinding solutions to keep those areas out of MG areas. Based on our pilot study, we can conclude that the new software is more practical and faster, shows an excellent correlation with the manual analysis, and has a high reproducibility.
REFERENCES
1. Knop E, Knop N, Millar T, et al. The international workshop on meibomian gland dysfunction: report of the subcommittee on anatomy, physiology, and pathophysiology of the meibomian gland. Invest Ophthalmol Vis Sci. 2011;52:1938–1978.
2. Nelson JD, Shimazaki J, Benitez-del-Castillo JM, et al. The international workshop on meibomian gland dysfunction: report of the definition and classification subcommittee. Invest Ophthalmol Vis Sci. 2011;52: 1930–1937.
3. Schaumberg DA, Nichols JJ, Papas EB, et al. The international workshop on meibomian gland dysfunction: report of the subcommittee on the epidemiology of, and associated risk factors for, MGD. Invest Oph-thalmol Vis Sci. 2011;52:1994–2005.
4. Mizoguchi S, Iwanishi H, Arita R, et al. Ocular surface inflammation impairs structure and function of meibomian gland. Exp Eye Res. 2017;163:78–84. 5. Arita R, Itoh K, Inoue K, et al. Noncontact infrared meibography to
document age-related changes of the meibomian glands in a normal population. Ophthalmology. 2008;115:911–915.
6. Arita R. Validity of noninvasive meibography systems: noncontact meibography equipped with a slit-lamp and a mobile pen-shaped meibograph. Cornea. 2013;32:S65–S70.
7. Arita R, Suehiro J, Haraguchi T, et al. Objective image analysis of the meibomian gland area. Br J Ophthalmol. 2014:746–755.
8. Dogan AS, Kosker M, Arslan N, et al. Interexaminer reliability of meibography: upper or lower eyelid? Eye Contact Lens. 2018;44:113–117. 9. Pult H, Nichols JJ. A review of meibography. Optom Vis Sci. 2012;89:
E760–E769.
10. Wise RJ, Sobel RK, Allen RC. Meibography: a review of techniques and technologies. Saudi J Ophthalmol. 2012;26:349–356.
11. Finis D, Ackermann P, Pischel N, et al. Evaluation of meibomian gland dysfunction and local distribution of meibomian gland atrophy by non-contact infrared meibography. Curr Eye Res. 2015;40:982–989. 12. Eom Y, Lee JS, Kang SY, et al. Correlation between quantitative
measurements of tearfilm lipid layer thickness and meibomian gland loss in patients with obstructive meibomian gland dysfunction and normal controls. Am J Ophthalmol. 2013;155:1104–1110.
13. Chen X, Utheim ØA, Xiao J, et al. Meibomian gland features in a Norwegian cohort of patients with primary Sjögren’ s syndrome. PLoS One. 2017;12: e0184284.
14. Yoo YS, Na KS, Byun YS, et al. Examination of gland dropout detected on infrared meibography by using optical coherence tomography meibography. Ocul Surf. 2017;15:130–138.
15. Wei A, Hong J, Sun X, et al. Evaluation of age-related changes in human palpebral conjunctiva and meibomian glands by in vivo confocal microscopy. Cornea. 2011;30:1007–1012.
16. Klee H, Allen R. Simulation of Dynamic Systems with MATLABand Simulink: Crc Press; 2018.
Manual versus semi-automated detector (operator I)
0.8383 ,0.001 Very good Manual versus semi-automated
detector (operator II)
0.7574 ,0.001 Good Semi-automated detector (operator I
vs. operator II)
0.9096 ,0.001 Very good Two measurements of
semi-automated detector operator I
0.8751 ,0.001 Very good
*Value of k and relevant strength of agreement according to Altman.
,0.20: poor; 0.21 to 0.40: fair; 0.41 to 0.60: moderate; 0.61 to 0.80: good; and.0.81: very good.
17. Hahn B, Valentine DT. Essential MATLAB for Engineers and Scientists. Cambridge, MA: Academic Press; 2016.
18. Flower MA. Webb’s Physics of Medical Imaging. Boca Raton, FL: CRC Press; 2012.
19. Pult H, Riede-Pult BH. Non-contact meibography: keep it simple but effective. Cont Lens Anterior Eye. 2012;35:77–80.
20. Pult H, Riede-Pult BH. Relation between upper and lower lids’ meibomian gland morphology, tearfilm, and dry eye. Optom Vis Sci. 2012;89:E310–E315.
21. Pult H, Riede-Pult B. Comparison of subjective grading and objective assessment in meibography. Cont Lens Anterior Eye. 2013;36:22–27.
22. Ban Y, Shimazaki-Den S, Tsubota K, et al. Morphological evaluation of meibomian glands using noncontact infrared meibography. Ocul Surf. 2013;11:47–53.
23. Koh YW, Celik T, Lee HK, et al. Detection of meibomian glands classification of meibography images. J Biomed Opt. 2012;17: 086008.
