The Impact of Age-Related Macular Degeneration on Retinal Layers Quantified by Deep Learning

Abstract

Purpose : To quantify the nonlinear relationship between retinal layers and age-related macular degeneration (AMD) with deep learning modeling.

Methods : From the population-based Leipzig Research Center for Civilization Diseases (LIFE) Adult Study, we included subjects with macular Spectralis optical coherence tomography (OCT) scans and AMD stage information. AMD was graded based on the Rotterdam classification method into three stages including non-AMD, early AMD (drusen size < 63 μm) and intermediate AMD (drusen size ≥ 125 μm). Late stage AMD participants were excluded from analysis. Ten retinal layers (Figure 1) segmented by the Heidelberg Engineering software were extracted including retinal nerve fiber layer (RNFL), ganglion cell layer (GCL), inner plexiform layer (IPL), inner nuclear layer (INL), outer plexiform layer (OPL), outer nuclear layer (ONL), myoid zone (MZ), ellipsoid zone (EZ) and outer-photoreceptor segment (OS) combined (EZ+OS), interdigitation zone (IZ) and retinal pigment epithelium (RPE). Ten retinal layers were used individually and together to predict AMD stages. The area under the receiver operating characteristic curve (AUC) was used to measure the association between retinal layers and AMD stages. Two-thirds and one-third of the data were used for training and testing with patient level separation, respectively.

Results : 15,623 OCT volume scans from 7,870 participants with AMD stages classified were included in this study (5,285, 2,449 and 136 subjects with non-AMD, early and intermediate AMD stages, respectively). The average age was 56.5 ± 12.3 years with 53.0% women. Combining 10 layers together, the AUCs to distinguish intermediate AMD from non-AMD, intermediate AMD from early AMD and early AMD from non-AMD were 0.90, 0.68 and 0.75, respectively. As expected, the outer layers of RPE (AUC: 0.87 and 0.73) and IZ (AUC: 0.87 and 0.73) were most strongly associated with early and intermediate AMD. Interestingly, ONL (AUC: 0.84 and 0.68) was associated with early and intermediate AMD with similar AUC strength as RPE and IZ.

Conclusions : We are able to quantify the nonlinear relationship between retinal layers and AMD stages with deep learning. ONL may encode important AMD information.

This abstract was presented at the 2023 ARVO Imaging in the Eye Conference, held in New Orleans, LA, April 21-22, 2023.