Purchase this article with an account.
Carla Agurto Rios, E Simon Barriga, Vinayak Joshi, Jeff Wigdahl, Cesar Carranza, Sheila C Nemeth, Wendall Bauman, Peter Soliz; Clinical Impact Of Image Quality Assessment In The Performance Of An Automated Diabetic Retinopathy Screening System. Invest. Ophthalmol. Vis. Sci. 2014;55(13):2285.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To assess the impact on clinical workflow and performance of an automatic image quality (IQ) process as part of a system for diabetic retinopathy (DR) screening in non-mydriatic fundus images.
In recent years several software systems for automatic IQ and DR assessment have been developed. This work assesses the clinical impact of image quality in automatic DR screening and proposes optimal IQ rejection rates that optimize performance of the DR screening system. The IQ software detects the presence of shadows and crescents, evaluates the overall image quality, and determines if the images are properly aligned. A case is considered to have sufficient quality for DR evaluation if at least one fovea-centered image from each eye is available. A case that does not comply with the requirement is considered incomplete. Only the images that pass the quality check are processed for DR. We varied threshold on the IQ algorithms to evaluate the impact on the performance of the DR screening system. For each threshold we calculated the number of inadequate cases which would need to be referred for re-imaging and the sensitivity and specificity of the DR screening system.
We tested the algorithms in 947 images (197 cases). For different IQ thresholds, total case rejection rates vary from 39% to 5%. As this rejection rate decreases, the total number of cases needed referral to retinal examination or re-imaging decreases from 61% to 56%. Based on these numbers the optimal operating point of the system would reject 17% of the cases while achieving a sensitivity of 94% for the detection of DR with a specificity of 67%.
The impact of IQ indicates a trade-off between the number of cases to be considered inadequate for grading by the algorithm and those incorrectly referred due to poor IQ. The optimal rejection percentage agrees with the 20% of inadequacy rate found in non-mydriatic screenings. Based in our data sample the amount of referrals can be reduced while still achieving high sensitivity for detection of DR. However, the specificity is also reduced.
This PDF is available to Subscribers Only