Purchase this article with an account.
Aaron S Coyner, Ryan Swan, Jayashree Kalpathy-Cramer, Sang Jin Kim, J. Peter Campbell, Karyn Elizabeth Jonas, Susan Ostmo, Robison Vernon Paul Chan, Michael F Chiang; Automated Image Quality Assessment for Fundus Images in Retinopathy of Prematurity. Invest. Ophthalmol. Vis. Sci. 2017;58(8):5550.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Accurate image-based ophthalmic diagnosis relies on clarity of fundus images. This has important implications for the quality of ophthalmic diagnosis, and for emerging methods such as telemedicine and computer-based image analysis. The purpose of this study is to implement an algorithm for automatically assessing the quality of fundus images, and to evaluate its performance on a set of retinopathy of prematurity (ROP) images compared to expert assessment.
A data set of 30 wide-angle ROP images (RetCam; Natus, Pleasanton, CA) was collected from infants undergoing routine screening examinations. A previously published method by Bartling et. al for automated assessment of digital fundus images was implemented. Briefly, 30 images were resized and converted to grayscale. Each image was broken into 16x16 pixel blocks, and blocks were assessed using metrics for brightness and sharpness. Scores for each block were used to compute an overall image quality score for each image. Images were ranked by quality score from 1 (lowest quality score) to 30 (highest quality score). Six independent experts were provided the same set of 30 images and, using a web-based interface, were prompted to "select the higher quality image for diagnosis of plus disease." Elo rating values for each image were determined for each rater and for the consensus of all raters. Spearman rank correlation was used to assess the similarity of the algorithm’s performance to human raters when determining image quality.
Individual expert ranks correlated highly with one another (correlation coefficient [CC] 0.89-0.94) and with the consensus rank (CC 0.94-0.98). The algorithm rank correlated with individual expert ranks (CC 0.81- 0.87) and the consensus rank (CC 0.86), but not as highly as individual experts (Table 1).
Experts in this study had high correlation to one another in assessing quality of ROP images. The computer-based algorithm assessed image quality with strong correlation to the expert consensus rank, although not as high as individual expert correlation with the consensus rank. Future studies will investigate discrepancies between the algorithm and the consensus, using the established consensus rank as a reliable training set.
This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.
Spearman's rank test correlation matrix of Consensus Rank, Algorithm Rank, and ranks from six individual experts.
This PDF is available to Subscribers Only