Abstract
Purpose :
In patients with intermediate age-related macular degeneration (AMD), the risk of progression to advanced stages is highly variable and the prognostic signs are unclear. We utilized machine learning to predict at which topographical location advanced AMD lesions will develop.
Methods :
OCT scans (512x128x1024 voxels, Cirrus) of fellow eyes of patients with choroidal neovascularization (CNV) undergoing regular monthly imaging were processed and registered at the time of advanced AMD onset as well as 1, 3, 6, 9 and 12 months prior. A deep learning network was employed to obtain automated segmentations of intraretinal fluid (IRF), subretinal fluid (SRF) and hyperreflective foci (HRF). In addition, automated segmentation of retinal layers and pigment epithelium detachment (PED) was performed. Quantitative features describing local retinal structure and appearance included inner retina, outer nuclear layer (ONL), photoreceptor outer segments with retinal pigment epithelium (RPE), and drusenoid space (Fig. 1). A total of 20 different topographic 2D feature maps were computed and rescaled to a 32x32 grid. A machine learning classifier was trained to predict for each grid element the local onset of advanced AMD.
Results :
The progression to CNV could be reliably diagnosed based on standardized evaluation of monthly OCT by two independent masked graders. The machine learning method was evaluated in a cross-validation setting on 86 eyes with PED and 27 without PED. The classification result of topographic predictions was compared to the reference standard CNV onset area defined by either PED or IRF. The area under the curve (AUC) for predicting the local AMD onset (Fig. 2) in eyes with PED was high (AUC=0.9) and the predictions relied largely on the drusen thickness maps. For the IRF, the predictions were more difficult (AUC=0.7), but were nevertheless feasible 3 months prior to the onset, with inner retinal layer and ONL properties found to have the most predictive role.
Conclusions :
Automated analysis of OCT biomarkers allows a personalized and localized prediction of AMD conversion. We proposed and evaluated a machine learning methodology to topographically predict the development of advanced AMD from OCT. The results of this pilot study are a promising step toward understanding the retinal characteristics and changes preceding the onset of advanced AMD
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.