Abstract :

Vision Loss From Retinal Disease Is Largely Preventable When Detected At An Early Stage; Yet Access To Expert Ophthalmological Screening Remains Constrained In Many Regions. We Introduce Retina Dx, A Hybrid Framework That Couples A DenseNet-121 Feature Extractor With A Novel Dynamic Spatial Graph Convolutional Network (DS-GCN). Unlike Prior Approaches That Rely On Fixed Or Hand-crafted Graph Topologies, Our DS-GCN Constructs, At Each Inference Step, A Per-image Adjacency Matrix Via Learnable Cosine-similarity Projections Applied To A 7×7 Spatial Patch Grid. A Top-K Edge-pruning Mask Then Suppresses Low-confidence Connections, And A Gated CrossAttention Fusion (GCAF) Module Adaptively Weights The CNN Global Descriptor And The GCN Relational Descriptor Before Classification. Trained On 10,133 Fundus Images And Evaluated On A Stratified Holdout Of 2,167 Images Spanning Seven Ocular Disease Classes, The Standalone Proposed DS-GCN+GCAF Hybrid Model Achieves 96.59% Accuracy (96.22% Precision, 97.96% Recall, 97.06% Macro F1-score), While The Full Multi-model Ensemble Achieves 98.71% Accuracy, 98.29% Precision, 99.21% Recall, And A Macro F1-score Of 98.74%. Three Layers Of Explainability Outputs Are Delivered Through A CPU-compatible Streamlit Interface, Enabling Clinician-facing Transparency Without GPU Dependency.

Published:

26-9-2026

Issue:

Vol. 26 No. 9 (2026)

Page Nos:

367-371

Section:

Articles

License:

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

How to Cite