Free
Cornea  |   August 2013
New Clinical Grading Scales and Objective Measurement for Conjunctival Injection
Author Affiliations & Notes
  • In Ki Park
    Department of Ophthalmology, Kyung Hee University College of Medicine, Kyung Hee University Hospital, Seoul, Korea
  • Yeoun Sook Chun
    Department of Ophthalmology, Chung-Ang University College of Medicine, Chung-Ang University Hospital, Seoul, Korea
  • Kwang Gi Kim
    Biomedical Engineering Branch, Division of Convergence Technology, National Cancer Center, Goyang, Korea
  • Hee Kyung Yang
    Department of Ophthalmology, Seoul National University College of Medicine, Seoul National University Bundang Hospital, Seongnam, Korea
  • Jeong-Min Hwang
    Department of Ophthalmology, Seoul National University College of Medicine, Seoul National University Bundang Hospital, Seongnam, Korea
  • Correspondence: Jeong-Min Hwang, Department of Ophthalmology, Seoul National University College of Medicine, Seoul National University Bundang Hospital, Gumiro 166, Gumi-dong, Bundang-gu, Seongnam, Gyeonggi-do, 463-707, Republic of Korea; hjm@snu.ac.kr
Investigative Ophthalmology & Visual Science August 2013, Vol.54, 5249-5257. doi:10.1167/iovs.12-10678
  • Views
  • PDF
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to Subscribers Only
      Sign In or Create an Account ×
    • Get Citation

      In Ki Park, Yeoun Sook Chun, Kwang Gi Kim, Hee Kyung Yang, Jeong-Min Hwang; New Clinical Grading Scales and Objective Measurement for Conjunctival Injection. Invest. Ophthalmol. Vis. Sci. 2013;54(8):5249-5257. doi: 10.1167/iovs.12-10678.

      Download citation file:


      © 2017 Association for Research in Vision and Ophthalmology.

      ×
  • Supplements
Abstract

Purpose.: To establish a new clinical grading scale and objective measurement method to evaluate conjunctival injection.

Methods.: Photographs of conjunctival injection with variable ocular diseases in 429 eyes were reviewed. Seventy-three images with concordance by three ophthalmologists were classified into a 4-step and 10-step subjective grading scale, and used as standard photographs. Each image was quantified in four ways: the relative magnitude of the redness component of each red-green-blue (RGB) pixel; two different algorithms based on the occupied area by blood vessels (K-means clustering with LAB color model and contrast-limited adaptive histogram equalization [CLAHE] algorithm); and the presence of blood vessel edges, based on the Canny edge-detection algorithm. Area under the receiver operating characteristic curves (AUCs) were calculated to summarize diagnostic accuracies of the four algorithms.

Results.: The RGB color model, K-means clustering with LAB color model, and CLAHE algorithm showed good correlation with the clinical 10-step grading scale (R = 0.741, 0.784, 0.919, respectively) and with the clinical 4-step grading scale (R = 0.645, 0.702, 0.838, respectively). The CLAHE method showed the largest AUC, best distinction power (P < 0.001, ANOVA, Bonferroni multiple comparison test), and high reproducibility (R = 0.996).

Conclusions.: CLAHE algorithm showed the best correlation with the 10-step and 4-step subjective clinical grading scales together with high distinction power and reproducibility. CLAHE algorithm can be a useful for method for assessment of conjunctival injection.

Introduction
Conjunctival injection or hyperemia is a nonspecific response with enlargement of conjunctival vessels induced by various diseases. Conjunctival injection is an important diagnostic clue for infection or inflammation and can be utilized for the monitoring of the disease progression and response to treatment. 1 However, the clinical grading of conjunctival injection is highly variable among clinicians. 2,3  
Various clinical classification methods for conjunctival injection have been introduced from the very simple binary scale (red or not red) to comparing the images of patients' eyes with reference images. 1,310 However, currently used photographs or illustration-based grading scales 11,12 —such as McMonnies/Chapman-Davies scale (MC-D), 7 Validated bulbar redness scale (VBR), 8 and Institute for Eye Research scale (IER)9—have some limitations, including unequal steps, biased description of references for actual levels of severity, restriction to specific conditions such as contact lens-related, and high intra- and interobserver variations, etc. 2,3,10,1215  
Clinical evaluation of conjunctival injection may be judged using two strategies. One is chromaticity and the other is vessel morphology, such as the area occupied by vessels, number of vessels, or diameter and tortuosity of vessels in the region of interest (ROI). The typical algorithm for color-based strategy is overall redness or color extraction using red-green-blue (RGB) systems, 2,16,17 and algorithms for the detection of morphologic properties of vessels are Canny edge detection 2,18 and the Hessian matrix, 19 which detect and extract the structural alterations at where the image brightness changes sharply or discontinues. During the last decade, those algorithms have been applied for the objective measurement of conjunctival injection but conflicting results have been shown in several studies. 2,16,17,20  
In this study, we tried to apply new strategies: the contrast-limited adaptive histogram equalization (CLAHE) and K-means clustering with LAB color model. CLAHE is a widely used technique for contrast enhancement of low density images. 21 It segments the original histogram into subhistograms with reference to brightness level. Consequently, the effect of variation of illumination is restricted to the local region, so overamplification is eliminated and characteristics of low-density histogram regions appear clearly. K-means clustering is a simple and well-known mathematical algorithm for grouping objects to handle large data sets. 22 Chromatic components of LAB color model 23 in each image were used as inputs of the K-means clustering algorithm to divide the image into three clusters: the red part (vessel region in conjunctiva), the white part, and the skin. In order to validate the diagnostic accuracy of these strategies, we needed to develop an ideal reference grading scale to overcome the limitations of the previously published standard photographs described above. Accordingly, the purpose of this study was to propose new clinical standard photographs and to determine the utility of the four objective assessment strategies for the grading of conjuncitval injection; the already well-known RGB and Canny-edge detection, newly applied CLAHE and K-means clustering with LAB color model. 
Materials and Methods
Grading Scale Image Collection
Four hundred and twenty-four anterior segment photographs with appropriate illumination, fine focus, high resolution, proper direction of fixation (maximum right, left, up, or down gaze to reveal a large area of conjunctiva) and view were selected from the database available at the Department of Ophthalmology, Seoul National University Bundang Hospital. The photographs were taken at ×10 magnification using a Haag-Streit BM 900 slit-lamp microscope (HAAG STREIT AG, Bern, Switzerland) in combination with a digital camera (Canon EOS 20D; Canon, Tokyo, Japan) interfaced to a personal computer, and was saved as a JPG file format (2544 × 1696 pixels, RGB, 16 MB). Slit beam lighting was used with maximum width (30 mm) and 4 mm height of the beam at a 45° oblique angle to take pictures of the bulbar conjunctiva. We used an automated digital camera system that set the aperture, shutter speed, and exposure time based on the external lighting conditions. Image acquisition, processing, and analysis were treated according to the tenets of the Declaration of Helsinki. This study was approved by the institutional review board (IRB). 
Three independent ophthalmologists with more than 15 years of clinical experiences were asked to grade the photographs from 1 (very mild) to 10 (very severe). Each clinician independently estimated the grade of photographs displayed on the same monitor, under identical room illumination without time limitation. Seventy-three photographs with concordant assessment by the three clinicians were sorted for a 10-step reference scale and image analysis; from grade 1 (mild) to 10 (very severe). Again, the three observers independently classified the photographs as grade 1 (mild); 2 (moderate); 3 (severe); and 4 (very severe) for simple and practical application. Among the 73 photographs, 69 photographs with concordant assessment between the three examiners were sorted. 
Region of Interest and Software Interface
An ROI is an area defined for further analysis or processing, leaving other regions unchanged. For objective assessment of conjunctival injection, we have to concentrate on a specific conjunctival region excluding the eyelid, cilia, cornea, or the superior or inferior tarsal conjunctiva. In this study, we used an interactive polygonal drawing method by clicking the left mouse button to begin, and moving the mouse along the border to finish with another left-click. This was repeated until the whole outline for the ROI was defined, displayed with a red color line on the monitor (Fig. 1B). 
Figure 1. 
 
Representative results of image analysis by Canny edge detection, K-means clustering with LAB color model, and CLAHE method. (A) Original image. (B) ROI. (C) Red conjunctival image detected by K-means clustering with LAB color method. (D) Image converted by gray scale and enhanced by CLAHE. (E) Thresholded image after enhancement. (F) Image of blood vessel edges detected by Canny edge detection algorithm.
Figure 1. 
 
Representative results of image analysis by Canny edge detection, K-means clustering with LAB color model, and CLAHE method. (A) Original image. (B) ROI. (C) Red conjunctival image detected by K-means clustering with LAB color method. (D) Image converted by gray scale and enhanced by CLAHE. (E) Thresholded image after enhancement. (F) Image of blood vessel edges detected by Canny edge detection algorithm.
The software shows the loaded picture file within its interface, and the ROI is drawn by the examiner. A few seconds later, the relative redness values of the ROI processed by each four algorithms are displayed on the left-lower side of the interface (Fig. 2). 
Figure 2. 
 
The interface of the novel software for objective measurement of conjunctival injection. After loading the picture file, the ROI is drawn by an interactive polygonal method, clicking the left mouse button to begin, and moving the mouse along the border to finish with another left-click. This is repeated until the whole outline for the ROI is defined. A few seconds later, the relative redness values of the ROI processed by the four algorithms are displayed on the left-lower side.
Figure 2. 
 
The interface of the novel software for objective measurement of conjunctival injection. After loading the picture file, the ROI is drawn by an interactive polygonal method, clicking the left mouse button to begin, and moving the mouse along the border to finish with another left-click. This is repeated until the whole outline for the ROI is defined. A few seconds later, the relative redness values of the ROI processed by the four algorithms are displayed on the left-lower side.
Objective Image Analysis
Each image was quantified by four algorithms: the relative magnitude of the redness component of each RGB pixels, the occupied area by the blood vessels based on K-means clustering algorithm and CLAHE, and the presence of blood vessel edges based on the Canny edge-detection algorithm (Fig. 1). Image analysis was performed using a newly developed software with personal computer system (Intel Core Quad CPU 2.40 GHz; Intel Corp., Santa Clara, CA). 
RGB Color Model.
In the RGB system, each pixel in the image is associated with three values corresponding to the intensities of the colors; red, green, and blue. The conjunctiva is perceived as red when the overall intensity of red (IR ) was higher than those of green (IG ) and blue (IB ). The level of redness can be considered proportional to the sum of (IR IG ) and (IR IB ). The relative redness for the ROI was calculated using the following equation2:  where S is the segmented ROI of the ocular region composed of n pixels and (IR)i , (IG)i , and (IB)i are the intensities of red, blue, and green components of the ith pixel in S, respectively. These values can range from −0.5 to 1. Pure red value is 1, white or gray is 0, and pure green or blue is −0.5.  
K-Means Clustering Algorithm With Lab Color Model.
LAB color model 23 is the most complete model used conventionally to describe all the colors visible to the human eye. This model was designed to be device independent, which means that the colors are defined independent of their nature of creation or the device they are displayed on. Therefore, its gamut exceeds those of the RGB color model, which represents only how the computer sees colors. LAB stands for Luminance (or lightness) of “A” and “B” (which are chromatic components). A ranges from green to red, and B ranges from blue to yellow. The inputted chromatic dataset is divided into three clusters and the data points are randomly assigned to the clusters with approximately the same number of data points. At each data point, the distance from the data point to each cluster is calculated. If the data point is closest to its own cluster, it is left where it is. If the data point is not closest to its own cluster, it moves onto the closest cluster. This process is repeated until no data point moves from one cluster to another. Consequently, three clusters (the red part, the white part, and the skin) are created through the color balancing of A and B by modifying output curves using the mathematical K-means clustering algorithm. 22 The red part of the conjunctiva was considered as the region of conjunctival vessels and it was calculated using the following equation:  where S is the segmented ROI of the ocular region and |S| is the total number of pixels in S. RegionRed Conjunctiva is the red conjunctiva cluster and |RegionRed Conjunctiva | is the total number of pixels in the red conjunctiva cluster. Therefore, V K-means (S) is the ratio of the number of pixels that are identified as redness to the total number of pixels in S, and these values range from 0 to 1 (Fig. 1C).  
CLAHE.
The original color image was first converted to a gray scale image and contrast was enhanced using CLAHE. While the general algorithm operates on the entire image, CLAHE partitions the image into small contextual regions (size of tiles: 8 × 8 with a clip limit of 0.01) and applies the histogram equalization to each one. By limiting the maximum and minimum ranges of equalization operations on individual tiles and by matching with multiple Gaussian transformation functions, the overequalization of the final image is eliminated. The enhanced image was thresholded and divided into red vessels and background. The feature based on this algorithm was calculated using the following equation:  where S is the segmented ROI of the ocular region and |S| is the total number of pixels in S. [CLAHE(S) > Threshold] i returns 1 if the ith pixel of S is bigger than the threshold value after enhancement with CLAHE; otherwise, it returns 0. Therefore, V CLAHE (S) is the ratio of the number of pixels that are identified as redness to the total number of pixels in S, and these values range from 0 to 1 (Figs. 1D, 1E).  
Canny Edge Detection.
To retrieve features related to the length of ocular blood vessels, an edge detection algorithm was adopted in this study. Because Canny edge detector 18 finds pixels on the edge of blood vessels, the number of detected pixels is proportional to the length of blood vessels regardless of variable widths along the structure of blood vessels. The feature based on the length of conjunctival blood vessels was calculated using the following equation:  where S is the segmented ROI of the ocular region and |S| is the total number of pixels in S. [Canny(S)] i returns 1 if the ith pixel of S is on the edge of blood vessel; otherwise, it returns 0. Therefore, V Canny (S) is the ratio of the number of pixels that are identified as edges to the total number of pixels of S, and these values range from 0 to 1 (Fig. 1F).  
Statistical Analysis
Conjunctival injection of all 73 photographs was objectively measured with four assessment strategies (RGB, K-means clustering with LAB color model, CLAHE, Canny edge detection). We evaluated the relationship between subjective grading scales and objective scores of the four algorithms using Spearman correlation. The scores of conjunctival injection measured by each four algorithms were compared using ANOVA. Significant difference was identified by ANOVA, post hoc analysis using the Bonferroni multiple comparison test. To determine the specificity and sensitivity of all four algorithms in differentiating the 4-step subjective clinical grading scale, receiver operating characteristic (ROC) curves and area under the receiver operating characteristic curves (AUC) were analyzed. To determine the degree of agreement between measurements conducted by two different observers (reproducibility), 73 photos were evaluated with CLAHE and their values were compared using ordinary least squares linear regression. Statistical analyses were performed using statistical software (SPSS software version 18.0, PASW, ver. 18.0; SPSS Inc., Chicago, IL). The α level (type I error) was set at 0.05. 
Results
Among the eligible 429 anterior segment photographs, 73 photographs (17%) with concordant assessment of three clinicians were sorted into a 10-step clinical grading system. Figure 3 shows the most representative 10 photographs. Among these 73 photographs, 69 photographs with concordant assessment of three clinicians were sorted into a 4-step clinical grading system. 
Figure 3. 
 
Representative photographs of standard images of conjunctival injection from grade 1 to 10, which were selected by three experienced ophthalmologists. The images labeled with an * represent the four-grade reference images.
Figure 3. 
 
Representative photographs of standard images of conjunctival injection from grade 1 to 10, which were selected by three experienced ophthalmologists. The images labeled with an * represent the four-grade reference images.
When compared with the 10-step subjective grading scale, three algorithms except Canny edge detection showed significant correlation. The occupied area by blood vessels based on CLAHE and K-means clustering algorithm showed better correlation than the relative magnitude of the redness by RGB. The strongest relationship was found with CLAHE (R = 0.919, P < 0.001). Canny edge detection based on the length of blood vessels showed no significant correlation (R = 0.094, P = 0.430; Table 1, Fig. 4). 
Figure 4. 
 
Graphs showing the relationship between the 10-step subjective clinical grading scales and redness scores measured by the four different objective methods. (A) Redness measured by RGB method. (B) K-means clustering algorithm with LAB color model. (C) CLAHE method. (D) Canny edge detection method. Mean and standard deviation bars are expressed.
Figure 4. 
 
Graphs showing the relationship between the 10-step subjective clinical grading scales and redness scores measured by the four different objective methods. (A) Redness measured by RGB method. (B) K-means clustering algorithm with LAB color model. (C) CLAHE method. (D) Canny edge detection method. Mean and standard deviation bars are expressed.
Table 1
 
Correlation Between Subjective Grading Systems and the Four Objective Measurements for Conjunctival Injection
Table 1
 
Correlation Between Subjective Grading Systems and the Four Objective Measurements for Conjunctival Injection
Measurement Correlation Coefficient, R P Value*
Subjective 10-grade system
 RGB system 0.741 <0.001
 K-means clustering with LAB 0.784 <0.001
 CLAHE 0.919 <0.001
 Canny edge detection 0.094 0.430
Subjective four-grade system
 RGB system 0.645 <0.001
 K-means clustering with LAB 0.702 <0.001
 CLAHE 0.838 <0.001
 Canny edge detection 0.092 0.451
When compared with the 4-step subjective grading scale, three algorithms except Canny edge detection showed significant correlation and the strongest relationship was found with CLAHE (R = 0.838, P < 0.001). Canny edge detection showed no significant correlation (R = 0.092, P = 0.451; Table 1, Fig. 5). The 10-step grading system showed stronger correlations with the automated techniques than the 4-step grading scale. 
Figure 5. 
 
Graphs showing the relationship between the four-step subjective clinical grading scales and redness scores measured by the four different objective methods. (A) Redness measured by RGB method. (B) K-means clustering algorithm with LAB color model. (C) CLAHE method. (D) Canny edge detection method. Mean and standard deviation bars are expressed.
Figure 5. 
 
Graphs showing the relationship between the four-step subjective clinical grading scales and redness scores measured by the four different objective methods. (A) Redness measured by RGB method. (B) K-means clustering algorithm with LAB color model. (C) CLAHE method. (D) Canny edge detection method. Mean and standard deviation bars are expressed.
In the 4-step grading scale, the mean measurement scores were significantly different between each grade in the three algorithms except Canny edge detection. Although RGB and K-means clustering with LAB partially differentiated the four grades, only CLAHE discriminated all four grades individually (Table 2). 
Table 2
 
Comparison of RGB, K-Means Clustering With LAB, and CLAHE With Subjective Four-Grade System for Conjunctival Injection
Table 2
 
Comparison of RGB, K-Means Clustering With LAB, and CLAHE With Subjective Four-Grade System for Conjunctival Injection
Algorithm Grade, A Grade, B Mean Differences, A–B SE P Value*
RGB system 1 2 −0.063 0.025 0.075
3 −0.109 0.058 0.002
4 −0.186 0.028 <0.001
2 3 −0.046 0.025 0.449
4 −0.123 0.025 <0.001
3 4 −0.077 0.029 0.055
K-means clustering with LAB 1 2 −0.058 0.017 0.006
3 −0.095 0.019 <0.001
4 −0.151 0.019 <0.001
2 3 −0.037 0.017 0.234
4 −0.092 0.017 <0.001
3 4 −0.056 0.020 0.039
CLAHE 1 2 −0.058 0.011 <0.001
3 −0.099 0.013 <0.001
4 −0.159 0.013 <0.001
2 3 −0.041 0.012 0.006
4 −0.101 0.012 <0.001
4 −0.060 0.013 <0.001
In the ROC curve of all four algorithms in differentiating the 4-step subjective clinical grading scale, AUC was largest for the CLAHE method, which was the only method with an AUC value of >0.9 for all grades. With the CLAHE method, the diagnosis of ≥grade 2 (cutoff 0.09); ≥grade 3 (cutoff 0.13); and ≥grade 4 (cutoff 0.16) showed sensitivities of 90.7%, 92.9%, and 92.9%, respectively, and specificities of 86.7%, 78.0%, and 80.5%, respectively (Fig. 6). 
Figure 6. 
 
ROC curves of all four algorithms in differentiating the four-step subjective clinical grading scales. AUC was largest for the CLAHE method, which was the only method with an AUC value of >0.9 for all grades. With the CLAHE method, the diagnosis of ≥grade 2 (cutoff 0.09); ≥grade 3 (cutoff 0.13); and ≥grade 4 (cutoff 0.16) showed sensitivities of 90.7%, 92.9%, and 92.9%, respectively, and specificities of 86.7%, 78.0%, and 80.5%, respectively.
Figure 6. 
 
ROC curves of all four algorithms in differentiating the four-step subjective clinical grading scales. AUC was largest for the CLAHE method, which was the only method with an AUC value of >0.9 for all grades. With the CLAHE method, the diagnosis of ≥grade 2 (cutoff 0.09); ≥grade 3 (cutoff 0.13); and ≥grade 4 (cutoff 0.16) showed sensitivities of 90.7%, 92.9%, and 92.9%, respectively, and specificities of 86.7%, 78.0%, and 80.5%, respectively.
The degree of agreement using CLAHE between two independent observers was very high (Y = 1.013X + 0.003, R = 0.996, R 2 = 0.992, P < 0.001 by linear regression). The reproducibilities of other algorithms were also good; RGB (Y = 0.989X + 0.009, R = 0.991, R 2 = 0.982, P < 0.001), K-means clustering with LAB (Y = 1.010X + 0.005, R = 0.993, R 2 = 0.987, P < 0.001), and Canny edge detection (Y = 0.991X + 0.025, R = 0.999, R 2 = 0.999, P < 0.001). 
Discussion
Newly applied objective assessment algorithms, CLAHE, and K-means clustering with LAB color models showed good correlation with the new clinical 10-step and 4-step grading scales. These models showed better correlations with subjective grades compared with the widely used RGB model. Among them, CLAHE method showed the best diagnostic accuracy with best distinction power and high reproducibility. 
To validate the utility of objective assessment methods, standard reference images for application of new systems were required. Currently, photographs or illustrations of reference grading scales such as McMonnies and Chapman-Davies, 7 Efron, 6 VBR, 8 and IER 9 are mostly oriented toward contact lens-related diseases. Among them, VBR scale was the latest psychophysically excellent grading system with proven physical characteristics. However, conjunctival injection was artificially induced by instillation of 5% hypertonic saline solution, and some images were modified using commercial graphics editing program (Adobe Photoshop; Adobe Systems, Mountain View, CA) to increase the amount of redness. Furthermore, subjective evaluations were performed by optometrists and optometry students, who may not take into account the various characteristics involved with conjunctival injection other than redness. 
In this study, we introduced new clinical standard reference scales (10-step and 4-step) with selected images originated from variable ocular disease conditions such as conjunctivitis, keratitis, uveitis, episcleritis, scleritis, pterygium, pingecula, side effects of eyedrops or contact lens. The clinical 10-step grading scale showed better correlation with objective measurement of conjunctival injection compared to the 4-step grading scale. Although conjunctival injection is a nonspecific response, its features are not alike. Even if images have similar degree of redness, the judgment by clinicians can be different according to their location, features, and types of disease. As shown in the report of Fieguth and Simpson, 2 the clinical judgments are highly variable and discrepancy reached up to 55% for each image. Cardona and Serés 24 reported that knowledge intensity and specificity influenced the grading skill and accuracy of conjunctival redness. Therefore, the integration of characteristics of conjunctival injection such as severity, location, morphology, and the experience of a clinician about conjunctival injection may have great effect on subjective assessment. Consequently, to improve the accuracy of subjective judgment, we selected concordant images that were originated from variable ocular conditions and evaluated by clinically experienced ophthalmologists. Just only 17% of eligible photos were finally chosen for the grading scales. In these respects, these new clinical reference scales have important advantages compared with other grading scales. 
Among the four objective assessment strategies, CLAHE and K-means clustering algorithm showed better correlations than RGB. Until recently, objective measurement of conjunctival injection was focused onto red color extraction and vessel edge detection, and there were not many papers discussing the importance of the occupied area by vessels in assessing conjunctiva injection. Wolffsohn and Purslow 17 emphasized the importance of color that red color extraction had the best correlation with IER subjective grade (R = 0.99), but the Canny edge detection showed a negative correlation (R = −0.81). Fieguth and Simpson 2 proposed a method for quantifying ocular redness by a combination of the overall redness and Canny edge detection algorithm. They showed excellent correlation with a 100-point subjective grading (R = 0.97), but the correlation was not linear and more discrepant for severe grades of conjunctival injection. Similarly, Peterson and Wolffsohn 20 demonstrated that the correlation with subjective grades was high (R = 0.98) when the combination of relative color extraction and vessel edge detection was applied. On the other hand, Papas 16 showed that the correlation with the number of vessels (R = 0.95) and the proportion of the image occupied by vessels (R = 0.96) was stronger than color-based algorithms (R = 0.7). However, this interesting result was drawn from almost normal bulbar conjunctival areas with relatively small regions (3.5 mm × 1.1 mm, 500 × 200 pixels), so it would differ from typical clinical gradings of conjunctival injection. 
Our study analyzed a relatively large region of bulbar area as shown in Figure 1 using CLAHE algorithm, which improves features of conjunctival blood vessels by enhancing the contrast of the applied image. The number of pixels representing vessel areas was divided by the total number of pixels in the binarized gray scale image to obtain the area occupied by vessels. A reasonable explanation of how the CLAHE estimates subjective grades more accurately than other techniques is thought to be the illumination-invariant characteristics of this algorithm. The overall redness is primarily a luminance–chromaticity-based judgment. If the redness increases, the luminance decreases. By this principle, the anterior segment images (especially, conjuncitval injection) captured by a slit-lamp biomicroscope feature a large variation in illumination. The CLAHE algorithm operates on small regions called tiles, and limits the ranges of equalization operations restricting the varying illumination. Therefore, the over- or underluminance equalization effect is eliminated and image contrast enhancement of conjunctival injection appears clearly. After Pizer et al. 21 introduced the CLAHE for medical imaging, it was applied to many medical fields such as portal films, 25,26 breast mammography, 27,28 ultrasound, 29 bone scan, 30 and fundus images for retinal hemorrhages detection. 31 Now, we can provide reasonable evidence that CLAHE can be used to assess the conjunctival injection objectively. The degree of agreement of CLAHE measurement between two independent observers was very high (slope = 1.0, R = 0.996). This represents not only the high reproducibility of the CLAHE method, but also the reliability of semiautomatic polygonal drawing of ROI. Hereby, we suggest that CLAHE is a simple, accurate, and effective measurement algorithm to evaluate the conjunctival injection objectively. 
Canny edge detection showed no significant association with subjective grades. The linear correlation was found only up to three steps in the 10-step grading scale, and up to two steps in the four-step grading scale. This may be attributed to its detection of pixels at the edge of blood vessels irrespective of its width. Therefore, it may only be proportional to the length or number of vessels and it could not provide information on vessel thickness and areas occupied by blood vessels. Moderate or severe conjunctival injection would be caused by increased vessels width and areas, not by increased vessel edges (number of vessels). This explains why the correlation between subjective and objective grades, evaluated by a combination of overall redness and Canny edge detection, was best only for low grades in Fieguth and Simpson's report. 2  
In summary, we introduced new clinical standard reference scales (10-step and 4-step) using images originated from variable and general ocular disease conditions. We covered the whole range of conditions that may cause conjunctival injection with various severity levels, not limited to specific conditions such as contact lens-related, etc. Newly applied CLAHE and K-means clustering with LAB showed better correlations with subjective grading than the widely used RGB system and Canny edge detection. This means that judgment of conjunctival injection by clinicians depend not only on red color but also the occupied area by vessels. 
In conclusion, CLAHE can be a useful algorithm for anterior segment image analysis of conjunctival injection independent of illumination. 
Acknowledgments
The authors alone are responsible for the content and writing of the paper. 
Disclosure: I.K. Park, None; Y.S. Chun, None; K.G. Kim, None; H.K. Yang, None; J.-M. Hwang, None 
References
Parrish RK Palmberg P Sheu WP. A comparison of latanoprost, bimatoprost, and travoprost in patients with elevated intraocular pressure: a 12-week, randomized, masked-evaluator multicenter study. Am J Ophthalmol . 2003; 135: 688–703. [CrossRef] [PubMed]
Fieguth P Simpson T. Automated measurement of bulbar redness. Invest Ophthalmol Vis Sci . 2002; 43: 340–347. [PubMed]
Bailey IL Bullimore MA Raasch TW Taylor HR. Clinical grading and the effects of scaling. Invest Ophthalmol Vis Sci . 1991; 32: 422–432. [PubMed]
Mandell RB. Slit lamp classification system. J Am Optom Assoc . 1987; 58: 198–201. [PubMed]
Hwang DG Schanzlin DJ Rotberg MH Foulks G Raizman MB. Levofloxacin Bacterial Conjunctivitis Place-controlled Study Group. A phase III, placebo controlled clinical trial of 0.5% levofloxacin ophthalmic solution for the treatment of bacterial conjunctivitis. Br J Ophthalmol . 2003; 87: 1004–1009. [CrossRef] [PubMed]
Efron N. Grading scales for contact lens complications. Ophthalmic Physiol Opt . 1998; 18: 182–186. [CrossRef] [PubMed]
McMonnies CW Chapman-Davies A. Assessment of conjunctival hyperemia in contact lens wearers. Part I. Am J Optom Physiol Opt . 1987; 64: 246–250. [CrossRef] [PubMed]
Schulze MM Jones DA Simpson TL. The development of validated bulbar redness grading scales. Optom Vis Sci . 2007; 84: 976–983. [CrossRef] [PubMed]
IER. IER grading scales. Institute for Eye Research: Sydney, Australia; 2007.
Chong E Simpson T Fonn D. The repeatability of discrete and continuous anterior segment grading scales. Optom Vis Sci . 2000; 77: 244–251. [CrossRef] [PubMed]
Schulze MM Hutchings N Simpson TL. Grading bulbar redness using cross-calibrated clinical grading scales. Invest Ophthalmol Vis Sci . 2011; 52: 5812–5817. [CrossRef] [PubMed]
Schulze MM Hutchings N Simpson TL. The use of fractal analysis and photometry to estimate the accuracy of bulbar redness grading scales. Invest Ophthalmol Vis Sci . 2008; 49: 1398–1406. [CrossRef] [PubMed]
Murphy PJ Lau JS Sim MM Woods RL. How red is a white eye? Clinical grading of normal conjunctival hyperaemia. Eye (Lond) . 2007; 21: 633–638. [PubMed]
Efron N Morgan PB Katsara SS. Validation of grading scales for contact lens complications. Ophthalmic Physiol Opt . 2001; 21: 17–29. [PubMed]
Peterson RC Wolffsohn JS. Sensitivity and reliability of objective image analysis compared to subjective grading of bulbar hyperaemia. Br J Ophthalmol . 2007; 91: 1464–1466. [CrossRef] [PubMed]
Papas EB. Key factors in the subjective and objective assessment of conjunctival erythema. Invest Ophthalmol Vis Sci . 2000; 41: 687–691. [PubMed]
Wolffsohn JS Purslow C. Clinical monitoring of ocular physiology using digital image analysis. Cont Lens Anterior Eye . 2003; 26: 27–35. [CrossRef] [PubMed]
Canny J. A computational approach to edge detection. IEEE Trans Pattern Anal Mach Intell . 1986; 8: 679–698. [CrossRef] [PubMed]
Sato Y Nakajima S Shiraga N Three-dimensional multi-scale line filter for segmentation and visualization of curvilinear structures in medical images. Med Image Anal . 1998; 2: 143–168. [CrossRef] [PubMed]
Peterson RC Wolffsohn JS. Objective grading of the anterior eye. Optom Vis Sci . 2009; 86: 273–278. [CrossRef] [PubMed]
Pizer SM Johnston RE Rogers DC Beard DV. Effective presentation of medical images on an electronic display station. Radiographics . 1987; 7: 1267–1274. [CrossRef] [PubMed]
MacQueen JB. Some Methods for Classification and Analysis of Multivariate Observations . Berkeley, CA: University of California Press. 1967; 281–297.
Schwarz MW Cowan WB Beatty JC. An experimental comparison of RGB, YIQ, LAB, HSV and opponent color models. ACM Trans Graph . 1987; 6: 123–158. [CrossRef]
Cardona G Serés C. Grading contact lens complications: the effect of knowledge on grading accuracy. Curr Eye Res . 2009; 34: 1074–1081. [CrossRef] [PubMed]
Rosenman J Roe CA Cromartie R Muller KE Pizer SM. Portal film enhancement: technique and clinical utility. Int J Radiat Oncol Biol Phys . 1993; 25: 333–338. [CrossRef] [PubMed]
Verellen D De Neve W Van den Heuvel F On-line portal imaging: image quality defining parameters for pelvic fields--a clinical evaluation. Int J Radiat Oncol Biol Phys . 1993; 27: 945–952. [CrossRef] [PubMed]
Pisano ED Zong S Hemminger BM Contrast limited adaptive histogram equalization image processing to improve the detection of simulated spiculations in dense mammograms. J Digit Imaging . 1998; 11: 193–200. [CrossRef] [PubMed]
Papadopoulos A Fotiadis DI Costaridou L. Improvement of microcalcification cluster detection in mammography utilizing image enhancement techniques. Comput Biol Med . 2008; 38: 1045–1055. [CrossRef] [PubMed]
Gómez W Leija L Alvarenga AV Infantosi AF Pereira WC. Computerized lesion segmentation of breast ultrasound based on marker-controlled watershed transformation. Med Phys . 2010; 37: 82–95. [CrossRef] [PubMed]
Jeong CB Kim KG Kim TS Kim SK. Comparison of image enhancement methods for the effective diagnosis in successive whole-body bone scans. J Digit Imaging . 2011; 24: 424–436. [CrossRef] [PubMed]
Bae JP Kim KG Kang JC Jeong CB A study on hemorrhage detection using hybrid method in fundus images. J Digit Imaging . 2011; 24: 394–404. [CrossRef] [PubMed]
Figure 1. 
 
Representative results of image analysis by Canny edge detection, K-means clustering with LAB color model, and CLAHE method. (A) Original image. (B) ROI. (C) Red conjunctival image detected by K-means clustering with LAB color method. (D) Image converted by gray scale and enhanced by CLAHE. (E) Thresholded image after enhancement. (F) Image of blood vessel edges detected by Canny edge detection algorithm.
Figure 1. 
 
Representative results of image analysis by Canny edge detection, K-means clustering with LAB color model, and CLAHE method. (A) Original image. (B) ROI. (C) Red conjunctival image detected by K-means clustering with LAB color method. (D) Image converted by gray scale and enhanced by CLAHE. (E) Thresholded image after enhancement. (F) Image of blood vessel edges detected by Canny edge detection algorithm.
Figure 2. 
 
The interface of the novel software for objective measurement of conjunctival injection. After loading the picture file, the ROI is drawn by an interactive polygonal method, clicking the left mouse button to begin, and moving the mouse along the border to finish with another left-click. This is repeated until the whole outline for the ROI is defined. A few seconds later, the relative redness values of the ROI processed by the four algorithms are displayed on the left-lower side.
Figure 2. 
 
The interface of the novel software for objective measurement of conjunctival injection. After loading the picture file, the ROI is drawn by an interactive polygonal method, clicking the left mouse button to begin, and moving the mouse along the border to finish with another left-click. This is repeated until the whole outline for the ROI is defined. A few seconds later, the relative redness values of the ROI processed by the four algorithms are displayed on the left-lower side.
Figure 3. 
 
Representative photographs of standard images of conjunctival injection from grade 1 to 10, which were selected by three experienced ophthalmologists. The images labeled with an * represent the four-grade reference images.
Figure 3. 
 
Representative photographs of standard images of conjunctival injection from grade 1 to 10, which were selected by three experienced ophthalmologists. The images labeled with an * represent the four-grade reference images.
Figure 4. 
 
Graphs showing the relationship between the 10-step subjective clinical grading scales and redness scores measured by the four different objective methods. (A) Redness measured by RGB method. (B) K-means clustering algorithm with LAB color model. (C) CLAHE method. (D) Canny edge detection method. Mean and standard deviation bars are expressed.
Figure 4. 
 
Graphs showing the relationship between the 10-step subjective clinical grading scales and redness scores measured by the four different objective methods. (A) Redness measured by RGB method. (B) K-means clustering algorithm with LAB color model. (C) CLAHE method. (D) Canny edge detection method. Mean and standard deviation bars are expressed.
Figure 5. 
 
Graphs showing the relationship between the four-step subjective clinical grading scales and redness scores measured by the four different objective methods. (A) Redness measured by RGB method. (B) K-means clustering algorithm with LAB color model. (C) CLAHE method. (D) Canny edge detection method. Mean and standard deviation bars are expressed.
Figure 5. 
 
Graphs showing the relationship between the four-step subjective clinical grading scales and redness scores measured by the four different objective methods. (A) Redness measured by RGB method. (B) K-means clustering algorithm with LAB color model. (C) CLAHE method. (D) Canny edge detection method. Mean and standard deviation bars are expressed.
Figure 6. 
 
ROC curves of all four algorithms in differentiating the four-step subjective clinical grading scales. AUC was largest for the CLAHE method, which was the only method with an AUC value of >0.9 for all grades. With the CLAHE method, the diagnosis of ≥grade 2 (cutoff 0.09); ≥grade 3 (cutoff 0.13); and ≥grade 4 (cutoff 0.16) showed sensitivities of 90.7%, 92.9%, and 92.9%, respectively, and specificities of 86.7%, 78.0%, and 80.5%, respectively.
Figure 6. 
 
ROC curves of all four algorithms in differentiating the four-step subjective clinical grading scales. AUC was largest for the CLAHE method, which was the only method with an AUC value of >0.9 for all grades. With the CLAHE method, the diagnosis of ≥grade 2 (cutoff 0.09); ≥grade 3 (cutoff 0.13); and ≥grade 4 (cutoff 0.16) showed sensitivities of 90.7%, 92.9%, and 92.9%, respectively, and specificities of 86.7%, 78.0%, and 80.5%, respectively.
Table 1
 
Correlation Between Subjective Grading Systems and the Four Objective Measurements for Conjunctival Injection
Table 1
 
Correlation Between Subjective Grading Systems and the Four Objective Measurements for Conjunctival Injection
Measurement Correlation Coefficient, R P Value*
Subjective 10-grade system
 RGB system 0.741 <0.001
 K-means clustering with LAB 0.784 <0.001
 CLAHE 0.919 <0.001
 Canny edge detection 0.094 0.430
Subjective four-grade system
 RGB system 0.645 <0.001
 K-means clustering with LAB 0.702 <0.001
 CLAHE 0.838 <0.001
 Canny edge detection 0.092 0.451
Table 2
 
Comparison of RGB, K-Means Clustering With LAB, and CLAHE With Subjective Four-Grade System for Conjunctival Injection
Table 2
 
Comparison of RGB, K-Means Clustering With LAB, and CLAHE With Subjective Four-Grade System for Conjunctival Injection
Algorithm Grade, A Grade, B Mean Differences, A–B SE P Value*
RGB system 1 2 −0.063 0.025 0.075
3 −0.109 0.058 0.002
4 −0.186 0.028 <0.001
2 3 −0.046 0.025 0.449
4 −0.123 0.025 <0.001
3 4 −0.077 0.029 0.055
K-means clustering with LAB 1 2 −0.058 0.017 0.006
3 −0.095 0.019 <0.001
4 −0.151 0.019 <0.001
2 3 −0.037 0.017 0.234
4 −0.092 0.017 <0.001
3 4 −0.056 0.020 0.039
CLAHE 1 2 −0.058 0.011 <0.001
3 −0.099 0.013 <0.001
4 −0.159 0.013 <0.001
2 3 −0.041 0.012 0.006
4 −0.101 0.012 <0.001
4 −0.060 0.013 <0.001
×
×

This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.

×