Purchase this article with an account.
Benton Chuter, Mark Christopher, Rui Fan, Akram Belghith, Christopher Bowd, Nicole Brye, James A Proudfoot, Jasmin Rezapour, Massimo Antonio Fazio, Michael Henry Goldbaum, Robert N Weinreb, Christopher A Girkin, Jeffrey M Liebmann, C Gustavo De Moraes, Linda M Zangwill; A deep learning model to assess fundus photograph image quality and improve predictive value of deep learning models of glaucoma detection.. Invest. Ophthalmol. Vis. Sci. 2021;62(8):1016.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To develop a deep learning model to appraise fundus photograph image quality, and to determine how the quality of photography influences the predictive value of separate glaucomatous optic neuropathy (GON) detection deep learning (DL) models.
A training dataset of 2,815 optics disc photographs acquired from healthy and GON patients as part of the Diagnostic Innovations in Glaucoma Study (DIGS) and African Descent and Glaucoma Evaluation Study (ADAGES) was used to develop a DL model to quantify the quality of photographs (0=good, 1=poor). Image quality ground truth was determined by two reviewers, with high quality images defined as sufficient to detect GON. The DL quality model was then applied to an independent test dataset of 11,350 photographs from the Ocular Hypertension Treatment Study (OHTS). A previously published DL model to detect GON in fundus photos was also applied to the OHTS data. To determine the impact of DL predicted quality on automated review of fundus photos, area under the receiver operating characteristic curve (AUC) was calculated for both good quality OHTS photographs (score 0.0-0.1) and poorer quality photographs (score > 0.1).
All OHTS photographs were considered good quality by the OHTS graders. In this dataset, 68 eyes reached OHTS primary open angle glaucoma (POAG) endpoint based of the development of visual field (41 eyes) or optic disc changes (55 eyes). The diagnostic accuracy of the DL model for POAG detection performed better in good quality compared to poorer quality photographs (AUROC: (95% CI)) of 0.86 (0.79, 0.91) and 0.81 (0.73, 0.88), respectively, though the differences did not reach statistical significance. Sensitivity at 90% specificity was also higher in the good quality (0.67) compared to poorer quality (0.44) photographs. Similar trends were found when the quality score was used to evaluate the DL glaucoma detection model separately on eyes that reached a visual field endpoint or an optic disc endpoint.
In the OHTS data, DL models for GON detection performed better in photos with DL predicted good quality, suggesting that DL based photograph quality assessment can be used to automatically identify low quality photos for removal. Incorporating quality assessment into automated review of fundus photos may thus improve GON detection model performance
This is a 2021 ARVO Annual Meeting abstract.
This PDF is available to Subscribers Only