Interpretable Retinal Disease Prediction Using Biology-Informed Heterogeneous Graph Representations

TL;DR

This work tackles DR staging from OCTA images with a focus on interpretability by introducing a biology-informed heterogeneous graph that encodes retinal vessels, intercapillary areas, and the FAZ. A graph neural network operates on this representation to perform DR staging, while an integrated explainability framework provides fine-grained spatial and feature-level attributions linked to clinically meaningful microvascular changes. The approach yields competitive ROC-AUC performance against strong CNN and transformer baselines, generalizes well to the OCTA-500 dataset, and delivers interpretable explanations that localize critical vessels and non-perfusion regions. The methodology offers a promising path toward clinically usable decision-support tools in ophthalmology by bridging high predictive accuracy with interpretable, anatomy-grounded reasoning.

Abstract

Interpretability is crucial to enhance trust in machine learning models for medical diagnostics. However, most state-of-the-art image classifiers based on neural networks are not interpretable. As a result, clinicians often resort to known biomarkers for diagnosis, although biomarker-based classification typically performs worse than large neural networks. This work proposes a method that surpasses the performance of established machine learning models while simultaneously improving prediction interpretability for diabetic retinopathy staging from optical coherence tomography angiography (OCTA) images. Our method is based on a novel biology-informed heterogeneous graph representation that models retinal vessel segments, intercapillary areas, and the foveal avascular zone (FAZ) in a human-interpretable way. This graph representation allows us to frame diabetic retinopathy staging as a graph-level classification task, which we solve using an efficient graph neural network. We benchmark our method against well-established baselines, including classical biomarker-based classifiers, convolutional neural networks (CNNs), and vision transformers. Our model outperforms all baselines on two datasets. Crucially, we use our biology-informed graph to provide explanations of unprecedented detail. Our approach surpasses existing methods in precisely localizing and identifying critical vessels or intercapillary areas. In addition, we give informative and human-interpretable attributions to critical characteristics. Our work contributes to the development of clinical decision-support tools in ophthalmology.

Figures (8)

• Figure 1: Schematic illustration of our heterogeneous graph representing the retinal vasculature. We elevate the image to a higher abstraction level, where nodes represent understandable biological concepts, such as vessels, intercapillary areas, and the FAZ. Because of each node's biological meaning, known tabular biomarkers are naturally encoded in our heterogeneous graph. Consequently, classifiers trained on our representation combine the favorable interpretability of tabular biomarkers with image-level localization and performance.
• Figure 2: Processing pipeline for generating a heterogeneous graph that models an OCTA image's most relevant biological concepts: vessels, intercapillary areas, and the FAZ. A vasculature segmentation kreitner2023detailed is used to create a vessel graph and intercapillary area graph and identify the FAZ. Finally, these components are merged into a single interconnected heterogeneous graph representation.
• Figure 3: Schematic illustration of our heterogenous graph representation of an OCTA image. Raw pixels in the image are abstracted into either a vessel node ($v_{ ves}$), an intercapillary area node ($v_{ ICA}$), or the FAZ ($v_{ FAZ}$). Neighborhood information is preserved in the homogeneous ($e_{ ves}$, $e_{ ICA}$) and heterogeneous edges ($e_{ ves-ICA}$, $e_{ F-ICA}$, $e_{ F-ves}$). Interpretable geometric descriptors and intensity statistics are encoded in the node embeddings.
• Figure 4: Illustration of the graph neural network architecture designed for the DR disease staging task. The architecture takes a heterogeneous graph as input and predicts the disease stage for the provided input graph. We generalize homogeneous message-passing architectures to the heterogeneous case by learning individual message-passing functions for each node type. After the message passing layers and linear layers, an aggregation for each node type is performed. Finally, the embeddings for all node types are aggregated and passed on to a classification head.
• Figure 5: Overview of cases where the microvascular changes assessed using en face OCTA images do not align with the disease progression information obtained from color fundus photography (CFP). (top row) Comparison of an en face OCTA image with the matching CFP image (middle and bottom row). Further cases of OCTA images where the OCTA-based expert assessment does not align with CFP grading.