Integration of Artificial Intelligence and OpacitySuppression^TM Software in Tele-Retinal Screenings

Abstract

Purpose : Images captured through non-mydriatic retinal cameras remain the standard in tele-retinal screening. Aging populations, with media opacity and small pupils, offer challenges in tele-retinal screenings that do not use mydriatic agents. Combining Artificial Intelligence (AI) software with image enhancement software such as OpacitySupression^TM (OS), to assist in interpreting retinal images in diabetic retinopathy screenings, has been evaluated.

Methods : 45° color retinal images were captured with a resolution of 21 Mp and a minimum pupil size of 2.8mm. 117 individuals with confirmed Type 1 Diabetic Mellitus (age 25 +/- 17.1, 42% male) were imaged having a natural pupillary dilation of >4.0mm. OS^TM was applied to images in ImageSpectrum (Canon, Los Angeles, CA), including 6 images (5%) that had been deemed unreadable by human readers due to pupil size bellow 2.8mm. OS^TM enhanced and OS^TM non-enhanced images were uploaded to a secure cloud and graded through AI software. XLSTAT was used to determine differences between the AI’s readings of enhanced versus non-enhanced images with respect to clinical screening parameters.

Results : AI read 117 images and flagged 49 positive pick-ups (42%); 38 microaneurysms, 6 exudates, and 5 hemorrhages. Paired T-test showed no significant difference between the AI’s detection with and without OS^TM for hemorrhages (mean pickup of 0.03 vs 0.04; p = 0.71), microaneurysms (0.32 vs 0.34; p = 0.68), exudates (0.05 vs 0.03; p = 0.26), and a difference in the ability to detect diabetic retinopathy (0. 49 vs 0.35; p = 0.03). All 6 images unreadable by a human reader were readable post OS^TM by a human reader, and OS^TM increased luminosity by an average of 34.7% in 6 images. While the AI was able to read these unreadable images without OS^TM, it did not flag any positive pickups. A human reader found a hemorrhage in one of the 6 previous unreadable images (17%) with the addition of OS^TM. The AI was not able to pick up any pathology in these images post-OS ^TM. AI interpreted images in 6 seconds while human reader’s interpretation time was 43 seconds per image.

Conclusions : Our results shows no significant difference in the ability of AI’s readings of enhanced versus non-enhanced OS^TM fundus images. However, OS^TM was found to be useful when human readers evaluate retinal images in community-based screenings.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.