A graph neural network-based clustering method for glaucoma detection from OCT scans considering uncertainties in the number of clusters

Abstract

Purpose : Clustering is desirable for unlabelled OCT scans which helps downstrean tasks like glaucoma detection and genetic analysis. However, determining the number of patterns is challenging and there is no ground truth in this regard. In this work, we propose an artificial intelligence model based on a graph neural network (GNN) that automatically determines the number of clusters and also clusters the images.

Methods : We first extracted the mean deviation (MD) of the visual tests within a time window of six months before and after for retinal nerve fiber layer thickness (RNFLT) images of Mass Eye and Ear dataset. Then we labeled the images in terms of being normal as normal (md>=0 dB) and non-normal (md<0 dB). An encoder (i.e., deep learning (DL) model) is trained with unlabeled images through unsupervised representation learning. The encoder then extracts the feature space of the image. Then, a graph neural network (GNN) is constructed by K nearest neighbor (KNN) with K=10, and trained with the features space of the labeled images (normal or non-normal) in which each image represents one node, and the feature space represents the node feature. During training, the graph is reconstructed iteratively based on the labels until it converges into a final graph (Figure 1). During inference, the graph takes the unlabeled feature space and clusters them into an unknown number of clusters.

Results : The GNN model was trained on 18507 scans for 5559 patients. Then, the trained GNN is tested on the unforeseen dataset of 5053 scans of 2064 patients. The GNN output 19 clusters for all scans, and filtering the normal predicted clusters resulted 6 clusters (patterns) remained for non-normal images (Figure 2A). The correlation with Pearson’s correlation coefficient between the features space of test images showed that while there is a strong correlation between all non-normal eyes, the correlation between the images within a cluster is stronger (Figure 2B).

Conclusions : The proof of concept of the proposed method resolves the challenge of uncertainty in the number of patterns in unlabeled OCT scans that help for glaucoma detection and analyses like the genome-wide association study (GWAS). In addition, due to the hierarchical learning of graph architecture, we can cluster OCT images that are taken from different devices with one trained GNN.

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