Deep learning-based detection of advanced AMD on retinal OCT from the UK Biobank dataset on behalf of the PINNACLE Consortium

Abstract

Purpose : To demonstrate the capability and limitations of artificial intelligence (AI) in automatically detecting late-stage age-related macular degeneration (AMD) in large non-disease-specific optical coherence tomography (OCT) volume data set from UK Biobank under population and image domain shift.

Methods : We applied a deep neural network, trained on a subset of independent real-world data from the retrospective part of the PINNACLE project (PIN), to the UK Biobank data set (UKBB). PIN consisted of 3,765 Topcon-2000 OCTs of 1,849 eyes with either AMD or a healthy retina. In contrast, UKBB consisted of 175,869 lower quality Topcon-1000 OCTs from 85,709 patients with a variety of diseases, where the majority is expected to be healthy, representative of population screening. The 3D volume classifier is composed of two convolutional neural networks, connected sequentially, and was trained to detect the presence of complete RPE, and outer retinal atrophy (cRORA) and/or macular neovascularization (MNV) on the 3D volume. Classification uncertainty estimates were generated with Monte-Carlo drop-out at inference time. After removing ungradable OCTs with a robustness filter, we classified 93,648 OCT out of 54,000 patients on UKBB. Uncertain predictions were removed to compensate for the scanner model domain shift. All remaining positive predictions for MNV and cRORA, 502 OCTs from 460 patients from UKBB, were graded by a retinal specialist. Furthermore, a hold-out test set of 96 OCTs from PIN was graded by a retinal specialist.

Results : Our AMD detection network achieved a positive predictive value (PPV) of 0.94 and 0.89 for MNV and cRORA on the test set of PIN, respectively. On the external UKBB, the classifier achieved a low PPV for the presence of active MNV (0.11) and a high PPV for cRORA (0.98). Reformulating the classification target to the presence of retinal fluid (nAMD, RVO, CSCR, DME...) the model achieved a PPV of 0.93 for retinal fluid, and 0.95 for fluid-and/or-atrophy vs no-fluid-and-no-atrophy on the UKBB.

Conclusions : Our results reflect the strengths and caveats of population-biased training of machine learning models. Although trained to detect MNV and cRORA secondary to AMD, the classifier was able to detect fluid and atrophy when exposed to a screening population beyond AMD eyes.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.