Graph Neural Networks Push Diabetic Retinopathy Detection Toward Near-Perfect Accuracy

Diabetic retinopathy, a progressive microvascular complication of diabetes that damages the light-sensitive tissue at the back of the eye, remains one of the leading causes of preventable vision loss worldwide. Caught early, it can be managed; missed, it can lead to irreversible impairment. Yet screening programs everywhere face the same bottleneck: there are far more retinal images to review than there are ophthalmologists to read them. A new study published in Discover Artificial Intelligence by S. Afzal, Biswambhar Rakshit and K. Somasundaram of Amrita Vishwa Vidyapeetham in Coimbatore, India, offers a fresh technical answer. Instead of relying solely on the convolutional neural networks that have dominated medical image analysis for a decade, the team built a hybrid framework that treats each retinal photograph as a graph, allowing a class of artificial intelligence known as graph neural networks to reason explicitly about the spatial relationships between lesions, vessels and healthy tissue.

The core insight behind the work is architectural. Convolutional neural networks such as ResNet and DenseNet are superb at detecting local patterns, textures and edges within small patches of an image, but they struggle to model the intricate topological structures and long-range spatial correlations that often distinguish a diseased retina from a healthy one. Diabetic retinopathy is precisely such a disease: microaneurysms, hemorrhages, exudates and abnormal vessel growth do not appear in isolation, and their distribution across the fundus carries diagnostic meaning. Graph neural networks, by contrast, operate on data represented as nodes connected by edges, and they learn by passing messages between neighboring nodes, iteratively aggregating information so that each node’s representation reflects both its own features and the context of its surroundings. That makes them naturally suited to capturing both local lesion detail and global structural organization in a way that fixed-grid convolutions cannot.

The researchers’ pipeline begins conventionally enough. They used the publicly available APTOS 2019 Blindness Detection dataset from Kaggle, consisting of fundus photographs resized to 224 by 224 pixels and Gaussian-filtered to reduce noise. The five original severity grades, ranging from no retinopathy to proliferative disease, were collapsed into a binary problem: any form of diabetic retinopathy versus none. Each image was then passed through a pre-trained ResNet50 backbone, with ImageNet weights providing a rich foundation of hierarchical visual features. Crucially, rather than collapsing the network’s output into a single global vector, the authors tapped an intermediate feature map of dimensions 512 by 28 by 28. Each of the 784 spatial positions in that map became a node in a graph, carrying a 512-dimensional feature vector that preserved the spatial structure of the underlying retina.

Edges between nodes were then formed using a k-nearest-neighbor strategy in feature space. For every node, the team computed Euclidean distances to all other nodes and connected it to its eight most similar counterparts, producing a graph in which proximity reflects feature similarity rather than mere pixel adjacency. The resulting graphs were split into training, validation and test sets of 70, 15 and 15 percent respectively, using stratified sampling to preserve class balance, and trained in mini-batches of 32 graphs on an NVIDIA Tesla P100 GPU. Two graph neural network architectures were trained and compared: GraphSAGE, which aggregates neighborhood information and concatenates it with the node’s own representation, and the Graph Isomorphism Network, or GIN, which applies a multilayer perceptron to the sum of a node’s features and those of its neighbors, giving it expressive power comparable to the Weisfeiler-Lehman graph isomorphism test.

Both models were trained for up to 100 epochs with the Adam optimizer at a learning rate of 0.001, using negative log likelihood loss and saving the weights that achieved the best validation accuracy. The results were striking. GIN reached an overall accuracy of 96.86 percent, with a recall of 98.66 percent and an F1 score of 96.96 percent, indicating an exceptional ability to catch true cases of the disease. GraphSAGE was close behind at 96.45 percent accuracy but led on precision at 95.53 percent, specificity at 95.29 percent and, most notably, an ROC-AUC of 0.992319 against GIN’s 0.986976, a near-perfect measure of the model’s confidence in separating the two classes. Both architectures comfortably outperformed state-of-the-art CNN baselines such as SE-ResNeXt50, EfficientNet and VGG-16, as well as a prior graph adversarial transfer learning approach that had reported 94.3 percent accuracy on the same dataset.

Perhaps the most scientifically interesting evidence came from the visualizations. When the researchers plotted the distributions of edge weights, the Euclidean distances between connected node features, they found that graphs from diseased retinas showed a sharp peak at low distance values, meaning pathological regions produce tightly clustered, locally coherent feature spaces. Healthy retinas, by contrast, yielded broader, more uniform distributions, reflecting greater textural and intensity variability. A principal component analysis of average node features reinforced the point: diseased and healthy graphs separated cleanly into distinct clusters in two-dimensional principal component space, demonstrating that the graph representations encode discriminative signals even before classification. In other words, the disease leaves a measurable structural fingerprint in the graph itself, something a standard CNN scanning local patches would not explicitly capture.

Interpretability, a persistent obstacle to clinical adoption of artificial intelligence, received dedicated attention. Because standard GIN and GraphSAGE models do not produce native attention coefficients, the authors computed gradient-based node saliency maps, scoring each node by the magnitude of the gradient of the predicted class output with respect to that node’s features. Nodes whose features most influence the prediction, and therefore correspond to regions the model relies on most heavily, light up in these maps, allowing clinicians to see which parts of the retina drove a diagnosis and to compare the structural patterns emphasized by each architecture. The team argues that such visualizations align with clinical observations and can build trust among medical professionals, a prerequisite for any diagnostic tool hoping to enter real workflows.

The study also probed robustness and statistical rigor. An ablation study trained the models on 10, 25, 50, 75 and 100 percent of the labeled data, with stratified sampling and multiple random seeds, to measure how performance degrades under scarce supervision, a common reality in medical imaging where expert annotations are expensive. A stratified five-fold cross-validation, with class weighting computed independently within each training fold and quadratic weighted kappa emphasized as the primary metric given the ordinal nature of the APTOS labels, confirmed the robustness of the results. A Wilcoxon signed-rank test on the paired performance differences between the two models produced a statistic of 7.0 and a p-value of 0.5625, meaning the observed differences between GIN and GraphSAGE were not statistically significant, so the choice between them can legitimately be guided by diagnostic priorities rather than raw superiority.

That trade-off has practical consequences. In a screening context, recall matters most, because a missed case of retinopathy, a false negative, can cost a patient their sight; GIN’s 98.66 percent recall makes it the natural choice when sensitivity is paramount. In a setting where false alarms trigger costly referrals and unnecessary anxiety, GraphSAGE’s higher precision and specificity become more valuable. The authors also analyzed computational complexity, showing that both architectures scale as O(L(Ed + Nd squared)) for L layers, E edges, N nodes and hidden dimension d, though GraphSAGE carries roughly twice the trainable parameters of GIN because its concatenation operation doubles the input dimensionality of each transformation layer. Notably, the entire pipeline, from graph construction through training and testing, ran in roughly sixteen minutes on a single cloud GPU, demonstrating that such models are feasible even in resource-constrained environments.

The researchers are candid that this is a step, not a destination. Future work, they write, should focus on optimizing hybrid CNN-GNN architectures that exploit the complementary strengths of both paradigms, expanding datasets with more diverse and longitudinal samples to improve generalizability, and streamlining the models for real-time clinical deployment. Integrating attention-driven and explainable artificial intelligence techniques, they argue, will be essential to earning clinician trust and embedding graph-based diagnostics into ophthalmology workflows. If the near-perfect discrimination reported here holds up in prospective, multi-center studies on unfiltered clinical data, graph neural networks could reshape automated retinal screening, catching the disease earlier, easing the burden on overworked specialists and, ultimately, preserving sight for millions of people living with diabetes.

Subject of Research: Graph neural network-based classification of diabetic retinopathy from retinal fundus images

Article Title: Graph hybrid neural network integration for enhanced diabetic retinopathy detection in retinal image

Article References: Afzal, S., Rakshit, B., & Somasundaram, K. (2026). Graph hybrid neural network integration for enhanced diabetic retinopathy detection in retinal image. Discover Artificial Intelligence, 6(1), Article 1427. https://doi.org/10.1007/s44163-026-02375-w

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02375-w

Keywords: diabetic retinopathy, graph neural networks, GraphSAGE, graph isomorphism network, deep learning, retinal images, medical image classification, APTOS 2019 dataset, ResNet50, computer vision, ophthalmology, explainable AI