Abstract
Purpose :
In this study, we evaluate the fully-automated DR screening system, EyeArt version 2.0, on a large (30,314) consecutive patient cases dataset from EyePACS telescreening setup. Current manual DR screening setups cannot scale to triage the ever-growing diabetic population at risk of vision loss. EyeArt can meet this growing need with automated, cloud-based screening.
Methods :
EyeArt’s hybrid approach combines novel morphological image analysis with state-of-the-art deep learning techniques to create a robust DR screening system engineered for large scale deployment on the cloud. EyeArt automatically analyzes multiple fundus images of a patient to generate patient-level DR screening recommendation, while flagging and excluding external eye images and poor quality images.
182,142 color retinal fundus images of 30,314 diabetic patient visits or encounters from the EyePACS setup were automatically analyzed by EyeArt 2.0 in under 15 hours. The encounters had 1-17 images including external eye images. EyePACS human graders provided the ICDR severity level and Clinically Significant Macular Edema (CSME) surrogate markers present/absent for use as reference standard to evaluate EyeArt’s DR screening recommendations. A patient was deemed non-referable if there was mild or no signs of DR and no CSME surrogate markers in both eyes
Results :
In the 30,314 encounters, 19.0% of encounters had moderate NPDR or higher or had surrogate markers for CSME and 5.2% of encounters had potentially treatable DR (severe NPDR or PDR). EyeArt’s screening sensitivity was 92.3% (95% CI: 91.6 - 93.0%) with specificity of 90.4% (95% CI: 90.0 - 90.8%). This corresponds to 7,488 “refer” cases and 438 false negatives out of which 93.2% did not meet the treatment criteria (i.e. had moderate NPDR). The AUROC was 0.958 (95% CI: 0.955 – 0.960). EyeArt’s sensitivity for detecting potentially treatable DR was 98.1%. 719 cases (2.5%) deemed non-screenable due to lack of at least 2 gradable images were given “refer” recommendations and included in results
Conclusions :
EyeArt 2.0 automated system has high screening sensitivity and specificity on a large real world dataset making it safe and effective for DR screening.
This is an abstract that was submitted for the 2016 ARVO Annual Meeting, held in Seattle, Wash., May 1-5, 2016.