Deep-learning-based clustering of OCT images for automated biomarker discovery in age-related macular degeneration

Abstract

Purpose : Current grading systems for age-related macular degeneration (AMD) have limited prognostic value for the conversion to late AMD. As such, there is an unmet need for imaging biomarkers that provide better stratification of patients. We explore using deep learning to automatically propose clusters of OCT images that share similar features, which we further investigate as candidates for new biomarkers.

Methods : Experiments were conducted using a dataset of 46,496 OCT images of AMD patients collected by the PINNACLE consortium. We first trained a ResNet50 neural network with the unsupervised contrastive learning method BYOL. The network learns, without any AMD labels, to extract disease-related features from OCT images, while ignoring trivial features such as image brightness or the orientation of the retina. The trained network then encodes each image in the lower-dimensional feature space (see Figure 1). Using k-means we divide the dataset among 30 clusters, assigning images with similar feature encodings to the same cluster. Finally, two expert ophthalmologists interpreted each cluster by reviewing 20 images from each, reporting any known, or potentially new, biomarkers. To aid this process, we highlight characteristic cluster features using GradCAM.

Results : Using our method we identified 30 clusters, some of which matched established biomarkers. For example, clusters 9, 27 and 16 separated cases with large drusen, cRORA (Complete RPE and Outer Retinal Atrophy) >250μm and cRORA >1000μm, respectively (see Figure 2). Two clusters, 18 and 22, divided exudative cases by those with subretinal vs. intraretinal fluid. Notably, some found biomarkers that have been linked to early AMD, such as serrated drusen (cluster 21), double-layer sign (cluster 14) and basal laminar deposits (cluster 10). Remarkably, one (cluster 24) displayed a potentially novel biomarker depicting cases of hypertransmission where the choroid remains thick and visible.

Conclusions : Clustering OCT images in large datasets, using deep learning, is a promising method for discovering new AMD biomarkers. Without any prior knowledge of AMD, our method rediscovered the set of biomarkers used in clinical grading systems and others that have previously been linked to early AMD. Most notably, we found one potentially new biomarker.

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

 

Workflow to propose new biomarkers using deep learning

 

Expert-derived descriptions of clusters