Grading static versus dynamic images of contact lens complications

Comparison of different smartphone cameras to evaluate conjunctival hyperaemia in normal subjects

Despite the significant advantages that smartphones’ cameras can provide in teleophthalmology and artificial intelligence applications, their use as black-box systems for clinical data acquisition, without adequate information of the quality of photographs can compromise data accuracy. The aim of this study is to compare the objective and subjective quantification of conjunctival redness in images obtained with calibrated and non-calibrated cameras, in different lighting conditions and optical magnifications. One hundred ninety-two pictures of the eye were taken in 4 subjects using 3 smartphone cameras{Bq, Iphone, Nexus}, 2 lighting levels{high 815 lx, low 122 lx} and 2 magnification levels{high 10x, low 6x}. Images were duplicated: one set was white balanced and color corrected (calibrated) and the other was left as it was. Each image was subjective and objectively evaluated. There were no significant differences in subjective evaluation in any of the conditions whereas many statistically significant main effects and interaction effects were shown for all the objective metrics. The clinician’s evaluation was not affected by different cameras, lighting conditions or optical magnifications, demonstrating the effectiveness of the human eye’s color constancy properties. However, calibration of a smartphone’s camera is essential when extracting objective data from images.

Assessing ocular bulbar redness: a comparison of methods

PURPOSE

We consider whether quantification of ocular bulbar redness, using image processing of relative Red-channel activity (Red-value), can be applied to a clinical sample and how this approach compares to an automated bulbar redness grading technique (Oculus Keratograph 5M, R-scan).

METHODS

Red-values from dry eye patients (n = 25) were determined using image processing of digital photographs over the nasal bulbar conjunctiva. Red-values were compared with subjective grades from six clinicians who graded the images using the IER scale. We considered the level of agreement between the Red-value and automated bulbar redness scores from the commercial instrument (R-scan). Scoring variability for each technique was assessed using the geometric coefficient of variation (gCoV, %). Agreement between techniques was considered with Bland-Altman analyses.

RESULTS

Red-values showed a strong linear relationship (R(2) = 0.99) to the R-scan. The Red-value had least variability (gCoV = 0.97%, 95% CI: 0.76-1.35%). The IER grade showed a linear relationship with Red-value (R(2) = 0.99), bound by a floor effect; it did not discriminate changes in redness below a threshold of 1.75 units (Red-value = 33.0%), after which it paralleled the redness returned by the R-scan. Intra-method variability for the redness returned by the R-scan (gCoV = 9.84%, 95% CI: 7.60-13.94%) and IER grades (gCoV = 7.30%, 95% CI: 1.73-10.31%) was similar (p > 0.05). Bland-Altman analysis showed the R-scan was consistently biased towards lower absolute redness scores than the IER.

CONCLUSIONS

Digital imaging processing, using relative Red-channel activity, was the least variable of the three techniques. The R-scan and IER showed similar intra-observer variability. The linear relationship between R-scan and Red-value suggests that the R-scan could be derived using similar methods.

Comparison between live and photographed slit lamp grading of corneal staining

PURPOSE

To compare live and photographic (still) grades of corneal staining of the same eyes and the repeatability of grading between two investigators.

METHODS

Thirty patients were recruited to participate in a contact lens study, and their level of corneal staining was graded by two investigators in situ (live images), using slit lamp biomicroscopy. Digital still images of the corneal staining were also captured during the study visits. An independent observer selected 105 of the still images graded by investigator 1 and another 105 images graded by investigator 2 and presented them to the original investigator in a random order, on three separate occasions. Grading was performed at the time of the live grading and the three still image sessions, using the Centre for Contact Lens Research corneal staining scale that combines grades of both extent and type to provide an overall "global staining score" from 0 to 10,000 for corneal staining. A comparison was made between live and still grades as well as the intrainvestigator repeatability for the multiple grading of the still images.

RESULTS

The mean (±SD) of corneal staining grades recorded for the same eyes examined live and then later on three occasions was 1795 (±1083) and 714 (±974), respectively, for participants examined by investigator 1 (p < 0.001) and 1854 (±1075) and 461 (±411) for those examined by investigator 2 (p < 0.001). There was a significant difference over the three repeated still grading sessions for each investigator (p < 0.001), although there was a high degree of consistency among the three still grading sessions for each of the investigators: the intraclass correlation for investigator 1 was 0.91 (confidence interval, 0.87 to 0.93) and that for investigator 2 was 0.82 (confidence interval, 0.77 to 0.87).

DISCUSSION

Digital still image grading of corneal staining significantly underrepresented the amount of corneal staining observed through a slit lamp. Clinical investigators graded corneal staining with a high degree of consistency.