Graph-Based Learning for Ophthalmic Image Analysis: Models, Methodologies, and Future Directions

Ophthalmic imaging serves as a cornerstone for the diagnosis and management of ocular diseases, capturing intricate anatomical structures and relational patterns that are not fully characterized by conventional convolutional neural networks operating on Euclidean grids. Graph-based learning offers a powerful paradigm to overcome this limitation by representing images as graphs, where nodes correspond to anatomical units and edges encode structural, spatial, or semantic relationships. This review provides a systematic and critical synthesis of graph-based methodologies applied to ophthalmic image analysis. It examines fundamental graph construction strategies-including node representation, edge formulation, and topological design-tailored to the hierarchical and relational nature of retinal structures. We further survey the evolution of graph learning models, from early graphical approaches to modern graph neural networks (GNNs) and their attention-based, diffusion-aware, and higher-order extensions. The discussion extends to learning strategies optimized for ophthalmic data challenges such as annotation scarcity, class imbalance, and domain shift. Clinically, we summarize representative applications across major imaging modalities, including fundus photography and optical coherence tomography, highlighting how graph-based frameworks advance tasks such as glaucoma assessment, diabetic retinopathy grading, vessel segmentation, and layered tissue analysis. Despite promising progress, critical challenges remain in robust and anatomically consistent graph construction, computational scalability, cross-domain generalization, and clinical interpretability. Future directions emphasize adaptive and uncertainty-aware graph building, scalable GNN architectures, integration of hypergraph representations for group-wise interactions, and unified multi-modal and longitudinal modeling. Through a structured analysis of models, methodologies, and applications, this review aims to guide the translation of graph-based learning into reliable and interpretable clinical tools for precision ophthalmology.