Purchase this article with an account.
Rory Sayres, Shawn Xu, T Saensuksopa, Marilyn Le, Dale R Webster; Assistance from a deep learning system improves diabetic retinopathy assessment in optometrists. Invest. Ophthalmol. Vis. Sci. 2019;60(9):1433.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Machine learning algorithms promise to improve diagnosis of retinal disease, through automated screening and/or augmenting clinicians. Previous research has demonstrated that providing algorithmic assistance can improve accuracy and grading confidence in ophthalmologists. However, optometrists may also be involved in identifying and referring diseases like diabetic retinopathy (DR). In this study, we examined whether assistance from a deep learning system (DLS) may similarly benefit optometrists.
We recruited 3 optometrists to read 399 cases each for DR severity using the ICDR scale. Readers graded 45° fundus images centered around the macula. Adjudicated grades from a panel of retina specialists served as the reference standard.Readers graded images in either an Unassisted or Assisted modality. Assignment of images to each condition was counterbalanced across images and readers. During the assisted condition, readers saw a predicted DR severity level produced by a DLS, and an explanatory heat map if Mild or worse DR was predicted. The heat map indicated regions indicative of any level of DR from Mild to PDR.
Without assistance, optometrist readers tended towards comparable sensitivity, but lower specificity, compared to the DLS. This is the opposite trend to that previously observed for ophthalmologists. Assistance tended to improve specificity without significantly affecting sensitivity for optometrists.Overall 5-class DR grading accuracy was not statistically different between arms (p = 0.18, logistic regression). However, 5-class accuracy was significantly greater for cases without any DR (p < 0.001), though not for cases with DR (p = 0.25).Assistance also improved optometrists’ self-reported grading confidence across all cases (p < 0.001, based on reader-reported confidence), and somewhat decreased the perceived difficulty of the grading task (p = 0.05, based on reader-reported difficulty). There were no significant effects of assistance on grading time (p = 0.84, linear regression).
Assistance from deep learning can potentially augment optometrists’ DR diagnosis performance and reduce the perceived cognitive burden of the task.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.
This PDF is available to Subscribers Only