Abstract
Purpose :
To investigate the relevance of an artificial intelligence (AI)-based diabetic retinopathy (DR) screening system coupled with a low-cost, hand-held fundus camera for meeting the diabetes care needs in a Hospital setting in Malawi, Africa.
Methods :
This investigation was a retrospective study of the performance of an AI software, EyeStar from VisionQuest Biomedical, to screen for referable DR using mydriatic images from a low-cost, hand-held, fundus camera, Pictor Plus, and secondarily using a smartphone-based mydriatic retinal camera, iNview, both from Volk Optical. Subjects were enrolled at the Lions Eye Unit, Zomba Central Hospital Zomba, Malawi, between December 2018 and March 2019.
The study enrolled N=122 subjects. Two images of the right eye were collected: one fovea centered and one optic disc centered. Eyes were pharmacologically dilated. All subjects were imaged with the Pictor camera and a subset of 36 with the iNview. The images were processed using the AI screening software which produces a “refer” or “non-refer” recommendation. The “refer” category corresponds to severe non-proliferative DR (NPDR) or proliferative DR (PDR). The “non-refer” category corresponds to no DR, mild NPDR, or moderate NPDR. The training of the AI software did not include images from the Pictor Plus. The performance of the AI software was assessed for sensitivity and specificity to detect referable DR using a certified reader as the reference standard. The Pictor test data included 88% with non-referable DR; and 12% with referable DR.
Results :
N=116 of the N=122 subjects had images from the Pictor Plus camera with sufficient quality for human reading to determine the level of DR. These 122 were presented to the AI software, which found 11 did not meet the quality requirements for an accurate prediction of DR grade. Without training on iNview images, the AI software processed successfully 20 cases. The Pictor achieved a sensitivity and specificity of 92% and 90% respectively, while the iNview achieved 100% and 76%.
Conclusions :
This study demonstrates the potential for the smartphone-based cameras to meet the needs in countries in terms of cost, iNview (<$1,000) and Pictor Plus ($10,000), and performance. With experience, the later stages of the study showed significantly reduced rejected images for quality.
This is a 2020 ARVO Annual Meeting abstract.