Background: We have observed that dry eye redness is characterized by a prominence of fine horizontal conjunctival vessels in the exposed ocular surface of the interpalpebral fissure, and have incorporated this feature into the grading of redness in clinical studies of dry eye. Aim: To develop an automated method of grading dry eye-associated ocular redness in order to expand on the clinical grading system currently used.
Methods: Ninety nine images from 26 dry eye subjects were evaluated by five graders using a 0–4 (in 0.5 increments) dry eye redness (Ora CalibraTM Dry Eye Redness Scale [OCDER]) scale. For the automated method, the Opencv computer vision library was used to develop software for calculating redness and horizontal conjunctival vessels (noted as “horizontality”). From original photograph, the region of interest (ROI) was selected manually using the open source ImageJ software. Total average redness intensity (Com-Red) was calculated as a single channel 8-bit image as R – 0.83G – 0.17B, where R, G and B were the respective intensities of the red, green and blue channels. The location of vessels was detected by normalizing the blue channel and selecting pixels with an intensity of less than 97% of the mean. The horizontal component (Com-Hor) was calculated by the first order Sobel derivative in the vertical direction and the score was calculated as the average blue channel image intensity of this vertical derivative. Pearson correlation coefficients, accuracy and concordance correlation coefficients (CCC) were calculated after regression and standardized regression of the dataset.
Results: The agreement (both Pearson’s and CCC) among investigators using the OCDER scale was 0.67, while the agreement of investigator to computer was 0.76. A multiple regression using both redness and horizontality improved the agreement CCC from 0.66 and 0.69 to 0.76, demonstrating the contribution of vessel geometry to the overall grade. Computer analysis of a given image has 100% repeatability and zero variability from session to session.
Conclusion: This objective means of grading ocular redness in a unified fashion has potential significance as a new clinical endpoint. In comparisons between computer and investigator, computer grading proved to be more reliable than another investigator using the OCDER scale. The best fitting model based on the present sample, and usable for future studies, was C CHOR CRED 4 2 2 = −12.24 + 2.12 + 0.88 :C4 is the predicted investigator grade, and CHOR 2 and CRED 2 are logarithmic transformations of the computer calculated parameters COM-Hor and COM-Red. Considering the superior repeatability, computer automated grading might be preferable to investigator grading in multicentered dry eye studies in which the subtle differences in redness incurred by treatment have been historically difficult to define.