Automated assessment of retinal image quality

Abstract

Abstract: : Purpose: Advances in digital imaging and computing have made automated quantitative analysis of retinal images feasible. The accuracy of image interpretation, however, depends in part on the quality of the images chosen for assessment. This study evaluates the performance of a prototype system to classify images according to multiple separate parameters of image quality. Methods: 530 color and fluorescein angiogram frames were graded (on a numeric scale of –2 to +2) by 3 examiners in 5 descriptive categories: exposure, lighting irregularity, contrast, focus, and artifacts. Images were randomly split into two groups (A and B) of equal size. Image parameters were calculated for each image including intensity statistics, first derivative estimations, and an approximation of the illumination image by a Gaussian blur with a kernel size equal to half of the image height. Pearson correlation coefficients were calculated between group A image parameters and subjective assessments of each descriptive category. A linear regression equation was calculated for the statistic in each category with the highest correlation and these regression lines were used to predict descriptive ratings for group B. Predicted values were compared to the subjective ratings for group B images using correlation coefficients. Results: The mean pixel intensity of the red channel correlated most strongly with exposure, the Std. Dev. of the illumination image intensity correlated best with lighting irregularity and the range (2 Std. Dev.) of the green channel intensity correlated with contrast (Table). Each of these parameters predicted group B ratings well. No parameters were found that predicted ratings in the focus or artifacts categories. 

Conclusions: Software algorithms can accurately assess several parameters which reflect retinal image quality. This ability may allow rapid automatic selection of optimal images for live feedback to photographers as well as preparation for quantitative analysis. These algorithms may be of value in systems being deployed for the automated screening of retinal disease (e.g. diabetic retinopathy) or for the quantification of complex diseases such as exudative AMD.

Keywords: imaging/image analysis: clinical • retina • image processing