Optic disc features in UKBIOBANK (UKBB) and the novel deep learning (DL) algorithm for automated prediction of cup-disc ratio (CDR) from colour fundus (CF) images.

Abstract

Purpose : To provide human grading results which describe features of suspicious glaucomatous disc damage on CF images in UKBB and preliminary results from the DL assisted image interpretation applied to these images.

Methods : Topcon 3D OCT–1000 Mark II system was used to acquire a single CF image of the optic disc and macula on un-dilated eyes. CF images were graded by NetwORCUK (a network comprising Belfast, Liverpool and Moorfields Ophthalmic Reading Centres) to answer pre-determined questions on features of the optic disc. Quality assurance was carried out at a rate of 1 in 20 gradings. CF images in the first batch graded, totaling 33222 participants, were cropped using the 'You Only Look Once' DL model to produce a sub-image of 299x299 pixels centred at the predicted centre of the optic disc before a novel end-to-end DL regression algorithm using an InceptionV3 based network was applied in order to predict CDR.

Results : 68517 participant image sets were analysed by graders for signs of glaucoma; 67985 right eye and 67613 left eye CF images were available. Of these 61,020 right eye and 60,242 left eye CF images were gradable.
CF images, regardless of quality, were included in the AI assisted image interpretation. High level of suspicion for glaucoma, defined as CDR >=0.7 was identified in 481 (0.70%) right eyes and 421 (0.61%) left eyes. There were an additional 299 (0.44%) of right eyes and 237 (0.35%) of left eyes with CDR < 0.7, with suspicious features such as haemorrhage on the disc or notching. In addition, there were 1210 participants with a difference of >=0.2 in CDR between eyes.
Preliminary AI results on the first batch graded using 62398 eyes (31788 right; 30610 left) allowed successful cropping in 54377 eyes (87%), of which 80% were used for training, 10% for validation and 10% for testing. The results on 5444 independent testing images showed the error between the predicted and true measurement of CDR to be <0.05 in 56% (95% CI: 54%, 57%), < 0.1 in 74% (95% CI: 76%, 78%) and <0.2 in 97.6% (95% CI: 97%, 98%) of images.

Conclusions : These data provide human grading for future projects on topics that require optic disc grading in both health and disease. AI assisted image interpretation has shown high accuracy when compared to human grading of the CDR on a large-scale.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.