Abstract
Purpose :
Previous deep learning (DL) attempts at predicting progression to late age-related macular degeneration (AMD) from color fundus photographs (CFP) have relied on ‘hand-crafted’ image features (drusen/pigmentary changes) and published 5-year risk estimates. The purpose was to develop, train, and validate a DL model for predicting progression from bilateral CFP using survival analysis and deep feature extraction. Deep feature extraction allows the unconstrained model to derive and weight multiple predictive image features, including complex features potentially absent from human grading systems.
Methods :
A DL survival (Cox proportional-hazards) model was trained to predict probability of progression to late AMD (neovascular AMD or central GA) at patient level for AREDS participants (55-80y, no AMD to non-advanced AMD at baseline). Four approaches were compared (see Figure). All models had demographic information as input. Based on bilateral CFP, Models 2-4 also received:
Model 2: grading for drusen/pigmentary changes by human retinal specialists;
Model 3: drusen/pigmentary changes predicted by DL (‘DeepSeeNet’, Peng et al., Ophthalmol. 2018);
Model 4: deep features extracted from DeepSeeNet fully-connected layer.
Each model was trained and tested on the same 3,535 and 710 participants, respectively (mean follow-up 8.2y). The primary outcome was concordance index (c-index, identical to AUC).
Results :
The c-index of each model was 0.703 (1), 0.718 (2), 0.884 (3), and 0.895 (4). The models using DL had substantially superior performance. Deep feature extraction (4) was superior to traditional features (3). In model 4, the most predictive characteristics were the imaging features (i.e. ranked above age/smoking).
Conclusions :
A DL model that combined deep feature extraction and survival analysis provided accurate time-based predictions of progression to late AMD. Deep feature extraction achieved superior performance to traditional features (assessed by retinal specialists or predicted by DL). By combining survival analysis with a deeper characterization of fundus images, this approach has advantages over previous attempts. This demonstrates the potential of deep feature-based biomarkers and survival analysis in improving AMD risk models.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.