Image processing techniques to quantify microprojections on outer corneal epithelial cells

Abstract

It is widely accepted that cellular microprojections (microvilli and/or microplicae) of the corneal surface are essential to maintain the functionality of the tissue. To date, the characterization of these vital structures has been made by analysing scanning or transmission electron microscopy images of the cornea by methods that are intrinsically subjective and imprecise (qualitative or semiquantitative methods). In the present study, numerical data concerning three microprojection features were obtained by an automated method and analysed to establish which of them showed less variability. We propose that the most stable microprojection characteristic would be a useful sign in early detection of epithelial damage or disease. With this aim, the scanning electron microscopy images of 220 corneal epithelial cells of nine rabbits were subjected to several image processing techniques to quantify microprojection density, microprojection average size and surface covered by microprojections (SCM). We then assessed the reliability of the methods used and performed a statistical analysis of the data. Our results show that the thresholding process, the basis of all image processing techniques used in this work, is highly reliable in separating microprojections from the rest of the cell membrane. Assessment of histogram information from thresholded images is a good method to quantify SCM. Amongst the three studied variables, SCM was the most stable (with a coefficient of variation of 15.24%), as 89.09% of the sample cells had SCM values ≥ 40%. We also found that the variability of SCM was mainly due to intercellular differences (the cell factor contribution represented 88.78% of the total variation in the analysed cell areas). Further studies are required to elucidate how healthy corneas maintain high SCM values.

Introduction

The flattened outer layer of squamous cells of the human cornea is covered by numerous fine microprojections (microvilli and/or microplicae) that enormously increase the surface area in contact with tears (Pfister, 1973). On this basis, several functions have been attributed to these characteristic structures. They may maximize the absorbance of oxygen and nutrients (Lemp et al. 1970; Beuerman & Pedroza, 1996; Ojeda et al. 2001) and stabilize the corneal tear film (Pfister, 1973; Harding et al. 1974). Moreover, microprojection functions seem to be closely related to the glycocalyx. Membrane-tethered mucins are placed on the tips of these protuberances, providing a lubricating surface that avoids microprojection damage during the blinking process (Gipson, 2004).

Although the specific role of corneal microprojections remains uncertain, it is widely accepted that they are essential to keep the physiology of the tissue intact (Pfister & Burstein, 1976; Collin & Collin, 2000; Johnson & Murphy, 2004). Consequently, correct assessment of the corneal surface must include a complete analysis of these membrane protuberances.

To our knowledge, no research has been conducted using quantitative, automated methods to evaluate microprojections. Additionally, current evaluation criteria and rating scales for corneal damage (Burstein, 1980; Berdy et al. 1992; Doughty, 2003) classify morphological changes undergone by microprojections into qualitative categories. This lack of objective results makes it difficult to identify conclusive features that underlie clinical and subclinical alterations in the corneal surface (Lemp, 1995).

Here we analyse different automated systems to quantify the morphological characteristics of corneal microprojections by means of several image processing techniques. In this way we can obtain numerical data about microprojection density, average size and surface covered by these cellular structures. We also show the outcomes of this quantitative evaluation on a sample of rabbit corneal epithelial cells and we establish which of these microprojection features is the most stable in healthy corneas.

Materials and methods

Two hundred and twenty cells of the central corneal epithelium of nine New Zealand white rabbits with normal eyes were selected for this study (mean of 26 cells per cornea). All animals were treated according to the ARVO Statement for the Use of Animals in Ophthalmic and Vision Research. The animals, weighing between 2 and 3 kg, were killed with pentobarbital sodium and immediately Karnovsky's fixative (Karnovsky, 1965) was dripped onto the corneal surface. Both eyes (right and left) were then enucleated and immersed in Karnovsky's fixative for 15 min. The corneas were then excised with a 1-mm scleral rim, transferred into a vial of fixative and kept in a refrigerator at 4 °C for 5 days. Corneas were then subjected to dehydration in a graded series of alcohols, critical-point dried, mounted on metal stubs with conductive silver paint and coated with gold.

The outer epithelial cells of the central zone of these specimens were examined using a Hitachi H2300 scanning electron microscope (SEM) at an acceleration voltage of 15 kV and ×5000 magnification. Cells that were totally or partially detached were excluded from the study because their microprojections must have undergone physiological degenerative processes (Doughty, 2003). Cell digital photomicrographs were taken using a personal computer with a frame grabber board as 8-bit greylevel images at TIFF resolution of 1024 × 832 pixels.

Visual and quantitative information about microprojections was obtained (see Table 1 for the description of variables). Numerical variables were calculated after applying two image processing techniques, namely binarization and particle analysis, which are explained in detail below. In both, Otsu's thresholding method (Otsu, 1979) was used.

Table 1.

Morphometric analysis of microprojections

Variable	Definition	Image processing	Quantifying method
Surface covered by microprojections (SCM)	Percentage of corneal area that is covered by microprojections.	Binarization	Histogram information
Microprojection density (MD)	Number of corneal microprojections per µm2.	Binarization + Watershed segmentation	Particle analysis
Surface covered by particles (SCP)	Percentage of corneal area that is covered by microprojections.	Binarization + Watershed segmentation	Particle analysis
Microprojection average size (MAS)	Microprojection average area (µm2)	Binarization + Watershed segmentation	Particle analysis

Finally, a statistical analysis of the data was made. We assessed the reliability of Otsu's method in this application. We compared the different methods and we also characterized the microprojections of healthy epithelial cells.

Binarization method

Image binarization or thresholding is a widely used method (González et al. 2004) that, according to an objective criterion, determines a grey threshold and assigns each pixel of a digital image to one class (image objects) if its grey value is greater than the determined threshold and otherwise to the other class (image background). An automated binarization was applied to a clean area (free of artefacts) of each selected cellular image (Fig. 1) by means of the free software ImageJ (W. Rasband, National Institutes of Health, Bethesda, MD, USA, http://rsb.info.nih.gov/ij). The automated thresholding function used by the ImageJ command Image/Adjust/Threshold applies Otsu's method (Otsu, 1979) (one of the most popular, simple and easy to implement), which is based on analysis of the shape of the image histogram. It establishes the threshold with an iterative process that computes the grey level average of the pixels at or below one initial threshold and pixels above. It then calculates the average of those two values, increases the threshold, and repeats the process. Iteration stops when the threshold is larger than the composite average. That is,

Fig. 1.

Process to obtain SCM values. With data from thresholded image histogram (number of black and white pixels) we obtained the percentage of pixels that represent microprojections.

Threshold = (grey level average of background + grey level average of objects)/2.

Therefore, by this method an image with different grey levels turns into a black and white one, separating the objects from the background. In our case, using binarization we were able to separate the pixels which represent cellular microprojections (brighter pixels) from those that represent the rest of the cell membrane (darker zones in an original image) (see Fig. 1 and following). Consequently, this allowed us to quantify the cellular surface covered by microprojections. We achieved this quantification on the basis of histogram information of thresholded images, i.e. the number of black and white pixels following binarization. According to these calculations, we defined a first variable called surface covered by microprojections (SCM), which is the percentage of pixels that represents these membrane protuberances over the total number of pixels in a cellular image (Fig. 1). Hence, SCM is the result of adding the area of all the microprojections in a cellular zone and calculating the percentage that it represents with regard to the total area analysed. We computed two SCM values (SCM₁and SCM₂) for all the sample cells, each one in a different zone of the cellular image. This allowed us not only to test the intercellular variability but also the variability in the same cell.

Moreover, we compared all the thresholded images with their respective references, the original images. This checking (González et al. 2004) allowed us to assess if binarization achieves our objective, i.e. successfully to separate microprojections from the rest of the cell membrane. For a complete comparison of the pairs of images (original and thresholded) we applied an arithmetic addition between them, pixel to pixel. This analysis shows whether automatic threshold generates underestimation (when it includes pixels that represent microprojections as background) or overestimation (when it includes pixels that represent cellular surface without microprojections as image objects) (see Figs 2 and 3). In case the resulting image from arithmetic addition displayed a conflict in demarcation of microprojections, we estimated the difference with histogram information.

Fig. 2.

Example of microprojection underestimation. You can see in (U₃) several grey microprojections that are not in the thresholded image (U₂).

Fig. 3.

Example of microprojection overestimation. You can see in (O₃) some parts of the intervening regions (darker areas) that have been included as microprojections in the thresholded image (O₂).

Particle analysis method

A second way to address the problem of microprojection quantification entails considering them as particles. In that case, after binarization (with Otsu's method; Otsu, 1979) we applied a watershed segmentation on the same selected area through ImageJ command Process/Binary/Watershed. This is a means of separating or cutting apart particles that touch, as corneal microprojections. This method finds ultimate erode points in the Euclidian distance map and dilates each of them as far as possible (either until the edge of the particle is reached or the edge of the region of another erode point) (see results in Fig. 4).

Fig. 4.

Watershed segmentation. (A) Particle analysis process in a cell with a high SCM. (B,C) The same process in a cell with medium and low SCM, respectively.

When microprojection images no longer overlapped, we counted them via a new ImageJ command (Analyze/Analyze particles). This works by scanning the image until it finds the edge of a particle, outlines the object and fills it. It then resumes scanning until it reaches the end of the image. For correct computing we configured the particle analyser in a size range of 9 pixels² to infinity, so the program ignores the particles outside this range. This minimal size restriction was made on the basis of a previous size study of the smallest microprojections of this sample in order to ignore some very small particles which were not microprojections and that we could see in several cell images. No restrictions were made for particle circularity. Finally, the ImageJ summary box showed us the number of particles included in this selection (see Microprojections). As we always selected an area of 10 µm² we could obtain a microprojection density measurement directly (particles/µm²) for each cell by a simple calculation. We named this new variable microprojection density (MD).

At the end of this process, the ImageJ summary box also showed a second variable that we called surface covered by particles (SCP), which represents the percentage of image area (pixels²) that we can consider particles, i.e. microprojections. This variable is conceptually the same as SCM (see Binarization above) although they are obtained in different ways, which allowed us to compare the two methods statistically in order to select one of them.

Furthermore, particle analysis provided data about a third variable, microprojection average size (MAS), through calculation of particle average area (in pixels²) (also shown in the ImageJ summary box). It is worth mentioning that in this case we excluded those particles that were on the edges of the selected zone because, obviously, their areas were not completely included within it. The measured values were converted to square micrometres by a simple arithmetic calculation; taking into account the magnification used (×5000).

Visual categorization

Apart from the calculation of quantitative variables we also made a visual evaluation of the surface covered by microprojections of each cell. In this way we added a new variable called SCMvisu (SCMv), which classified the cells into three categories: high (1) (the distance between two microprojections was, usually, similar to microprojection dimensions), medium (0.5) (the distance between two microprojections was, usually, longer than microprojection dimensions) and low (0.2) (in case of isolated microprojections) (see examples in Fig. 4). Estimates were masked to any knowledge of other measurements. Moreover, SCMv values were compared with SCM values in order to evaluate their degree of resemblance.

Results

Reliability assesment of thresholding images

The first step of the two automated quantitative methods used in this work (binarization and particle analysis) was to apply Otsu's thresholding criterion (Otsu, 1979) in a cell image to obtain a selection of its pixels. None of the 220 thresholded images included in this work showed an overestimation in separating pixels that represented microprojections from those that represented the rest of the cell membrane (the overestimation example of Fig. 3 is due to the application of a forced threshold) and only 18 images (8.18%) exhibited underestimation. The disagreement mean value was estimated at less than 4% from the total pixels analysed in an image.

It is also important to note that all the original images of the group with underestimation showed microprojections with a higher greylevel variation. This heterogeneity generated a three-peaked image histogram (see Fig. 5), instead of the typical two-peaked one (one peak for grey levels of microprojections and another for cellular surface without microprojections). It is well known that in these cases Otsu's method is less effective in distinguishing objects from the image background (Otsu, 1979). We also verified that this microprojection variability always appeared in small groups of adjacent cells. It could be the result of stress from sample manipulation, as other authors have described (Pfister & Burstein, 1976; Doughty, 1990). Moreover, one of these cell images, which showed an unusually high disagreement (12.13%), was slightly out of focus (see Fig. 5).

Fig. 5.

Examples during reliability assesment of thresholding images. (A) An example of agreement between the original image and its thresholded one. (B) A typical subestimation (1.84%). (C) The image with the highest subestimation in the sample (12.13%). It is evident that a small focusing failure exacerbates the three-peaked shape of the histogram (C₂). (Note that there is one peak at the level 0 and another at the level 255 in all the histograms).

Comparison of automatic methods: SCM versus SCP

For assessing agreement between the two methods that quantify the percentage area covered by microprojections we applied a Bland and Altman's statistical analysis (Bland & Altman, 1986) to SCM and SCP. A plot of the percentage difference versus mean for each pair of values is shown in Fig. 6. Agreement between the two methods decreased as the magnitude of SCM and SCP increased, such that higher differences corresponded to cells with a high surface area covered by microprojections. Furthermore, the variability of the results (Fig. 6) showed a slight tendency to increase with mean value.

Fig. 6.

Scatter plot of difference in percentage scale versus mean for each pair of SCM and SCP values. The solid line shows the average difference, and the broken lines show the 95% limits of agreement.

SCM vs SCMv

The relationship between the calculated values of surface covered by microprojections (SCM) and its visual assessment (SCMv) is displayed in Fig. 7 as a multiple box-plot. It assigns SCM values to SCMv categories. There was good agreement between higher and lower SCM values with their corresponding SCMv categories. Nevertheless medium values of SCM (35–45%) can be classified visually as either medium (0.5) or high (1) surface area covered by microprojections. Difficulties in visual classification of this group were mainly due to the fact that epithelium showed two different ways for keeping a surface covered to medium levels. The first was through a high MAS in cells with a relatively low density of microprojections. The second one was through a relatively high microprojection density in cells with a lower MAS. We found that in the SCM range 35–45%, cells with a medium SCMv score had MAS values of between 0.025 and 0.05 µm² whereas cells with a high SCMv score had MAS values of between 0.015 and 0.024 µm². Despite the overlap, there was a statistically significant difference between each pair of means corresponding to the three categories (P < 0.0001).

Fig. 7.

Quantitative (SCM, %) vs visual estimation (SCMv categories) of surface covered by microprojections.

Characteristics of microprojections in healthy corneas

The general appearance of healthy epithelium in the sample corneas agrees with the typical description that appears in the literature (Pfister, 1973; Doughty, 1990; Beuerman & Pedroza, 1996; Collin & Collin, 2000). According to this, at ×500 magnification, the tissue showed an uninterrupted cell mosaic with different grey tonalities, shapes and sizes. The cell nuclei were not evident except in very few cells in the process of desquamation (which were excluded from the present study). Almost all the cells were covered by a high number of microprojections that displayed a high individual uniformity at ×5000 magnification but a considerable variety of shapes amongst different cells (e.g. compare the images included herein).

For morphometric characterization of these membrane structures we calculated three different variables (see Table 1) in our sample for quantification: surface area covered by microprojections, microprojection density and average size of microprojections (SCM, MD and MAS, respectively). Their summary statistics (see Table 2) showed that SCM had the narrowest range (CV 15.24%) and its values were generally high (89.09% of the sample cells have SCM values ≥ 40%).

Table 2.

Summary statistics of the sample (SCM, %; MD, microprojections per µm²; and MAS, µm²)

	Mean	SD	CV (%)	Min.	Max.
SCM	48.28	7.36	15.24	17.58	60.61
MD	17.45	4.71	27.02	5.70	29.80
MAS	28.54 × 10−3	85.30 × 10−3	29.89	14.65 × 10−3	55.29 × 10−3

Analysing the pairwise correlation amongst these three variables (SCM, MD and MAS) only the MD vs MAS had a high coefficient in absolute value (–0.858, P < 0.01) although all the correlations were significant (SCM vs MD, P < 0.01; SCM vs MAS, P < 0.05). The conceptual equivalence between SCM (%) and MD*MAS*100 is shown in Fig. 8. The regression line (r = 0.96) is very close to the diagonal, which would represent an exact relationship, thus revealing the high degree of accuracy in the calculation of these microprojection features.

Fig. 8.

Linear relationship between surface covered by microprojections (SCM, %) and microprojection density (MD) multiplied by microprojection average size (MAS)*100.

As SCM seems to be the most stable and complete variable, we evaluated its sources of variability by applying a variance component analysis. Four factors account for the variability of the surface area covered by microprojections: the rabbits selected for study, the corneas corresponding to these rabbits (left or right), the selected cells on the corneal epithelium and finally the examined zones on each cell (1, 2). This analysis of variance (see Table 3) divided the total variability into four components corresponding to the four factors, each factor nested in the one above. The source contributing most to the variance was the variability in SCM over the cells included in the sample. Its contribution represented 88.78% of the total variation in the cell areas analysed. Variability contributed by the zones of the cells analysed was quite low, 2.54%, showing the relative uniformity of the surface covered by microprojections in the same cell.

Table 3.

Factors that account for the variability of SCM

Source	Sum of squares	d.f.	Mean square	Variance components	Percentage
Total	24021.9	439
Rabbit	1412.42	4	353.10	0.97	1.74
Cornea	1206.52	4	301.63	3.85	6.94
Cell	21092.7	211	99.96	49.28	88.78
Zone	310.20	220	1.41	1.41	2.54

Discussion

The reliability results of thresholded images confirm that binarization achieved our objective, i.e. successfully separated microprojections from the rest of the cell membrane. This is due mainly to two factors. First, the scanning electron microscope has a high resolution and a large depth of focus that produces high-quality images with little noise. This is why image focus is so decisive for producing a successful result. Second, almost all the cell images showed a homogeneity of microprojection grey distribution that generates a two-peaked histogram (one peak for the microprojection grey-level distribution and another for the intervening regions) and it is in these cases when Otsu's criterion is optimal (González et al. 2004; Li et al. 2005).

The disagreement found between the two methods for quantifying the percentage area covered by microprojections (binarization or particle analysis) appears to be caused by the loss of microprojection area entailed in the segmentation process via particle analysis (in SCP calculation). This effect is more notable as the density of microprojections increases (see Fig. 4) because membrane prominences have a greater tendency to touch each other. Moreover, the degree of junction depends on the microprojection shape, which is quite variable particularly at high SCM values (compare Figs 1 and 4). This could explain the higher variability in these cases.

Therefore, it seems evident that the best method to quantify SCM is the simplest, i.e. using histogram information of thresholded images. In fact, SCM calculation is also better than its visual evaluation (SCMv), as would be expected, as SCM is more discriminatory and more objective than SCMv. Additionally, visual assessment of microprojection characteristics is a difficult task, even for an expert, in view of the complexity of microprojection arrangements. As we have shown here, the visual method is an adequate system if the only goal is to identify cells with high or low SCM value. However, this is a poor method if we wish to detect early epithelial changes.

Our goal was to use an automated system to quantify microprojection features in order to establish which are the most stable in healthy corneas. Previous authors have measured microprojection height, width and density manually (Blümcke & Morgenroth, 1967; Pfister, 1973). Instead, we used an automated method to measure microprojection density, average size and surface covered by microprojections. Our results concerning microprojection density are similar to those fro the above studies although our range of data was wider, perhaps because we included many more cells in the sample.

With regard to average size (MAS) and surface covered by microprojections (SCM), to our knowledge, this is the first time that they have been measured. It is worth mentioning that both depend, via SEM imaging, not only on the dimensions of these membrane protuberances but also on their arrangement on the corneal surface (SCM depends on their density as well).

Amongst the three microprojection characteristics assessed herein, surface covered by these structures is the most stable in healthy corneas. This finding supports the idea that the functions of membrane protuberances are closely connected with their surface (Lemp et al. 1970; Beuerman & Pedroza, 1996; Ojeda et al. 2001). Because of the stability of SCM, we consider that this would be a useful measure to detect early epithelial damage. Further studies are required to elucidate this question as well as the physiological mechanism that maintains a relatively constant surface covered by microprojections in the healthy corneal epithelium.

Moreover, it is quite possible that this mechanism will only be effective when glycocalyx is intact as the lack of this structure has been linked to pathological changes in microprojections (Koufakis et al. 2006). This leads us to suggest that the membrane-tethered mucins (Argüeso & Gipson, 2001; Argüeso et al. 2003) could also have a role in the establishment of the microprojection arrangement pattern.

By contrast, the variability in SCM is mainly due to intercellular differences whereas this microprojection characteristic is relatively stable on individual cells. These data agree with morphological descriptions (Pfister, 1973; Harding et al. 1974; Doughty, 1990; Beuerman & Pedroza, 1996) and open the door to new investigations. In the near future, our quantification methods will allow us to assess the variation of SCM in the classical three groups of epithelial cells (light, medium and dark electron reflexes by SEM). This knowledge, as well as the study of this variable in adverse corneal conditions, will provide more accuracy in establishing the normal range of values of SCM in corneal disease.