Abstract
Purpose :
The retina's vasculature intrinsically forms a physical graph structure. Advanced concepts from geometric deep learning, such as graph convolutional networks (GCN), can solve complex machine learning tasks on any graph. As such, geometric deep learning, specifically graph learning lends itself excellently to advancing the automated analysis of the retinal vessels and associated diseases. In this study, we demonstrate the feasibility of vascular graph representations to ophthalmology by using a GCN to diagnose diabetic retinopathy (DR) from OCTA images.
Methods :
Our method consists of three steps. First, we apply a U-Net, trained on synthetic OCTA images (Menten et al., 2021), to segment the vessels. Second, we extract the vascular graph, where vessel bifurcations are the nodes, and the connecting vessels are the edges using the Voreen framework. Third, we apply a GCN to classify the vascular graphs from the OCTA images into healthy volunteers and DR patients. The GCN consist of three GC layers with 64 hidden channels. We use global mean pooling and train using the weighted binary cross entropy loss with a learning rate of 0.005 and weight decay of 5e-5. The final classification layer is a single linear layer.
We conduct our experiments on the public OCTA-500 dataset. We use 160 OCT images of healthy patients and 29 DR patients. We select our model after 20 epochs of training and carry out five-fold cross-validation on an 80/20 data split.
Results :
Our experiments show that our GCN is capable of classifying DR in vascular graph representations extracted from OCTA. We achieve an excellent accuracy of 0.95±0.03 and a balanced accuracy (Bal. Acc.) of 0.88±0.10 across all cross-validation folds. We further find that our GCNs outperform traditional baselines such as support vector machines (Bal. Acc. 0.81±0.02) and random forest models (Bal. Acc. 0.87±0.03)
Conclusions :
GCNs have proven to be powerful and efficient neural networks for diverse applications. The retinal vasculature forms a natural graph, which facilitates our application of GCNs to classify DR in retinal vessel graphs, which are extracted from OCTA images. Our experiments show that GCNs are highly accurate and outperform other machine learning baselines in DR detection, potentially indicating a path toward the application of GCNs in OCTA image analysis.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.