Journal: bioRxiv
Article Title: A graph-based deep learning framework for diabetic retinopathy classification with topology-aware feature augmentation
doi: 10.64898/2026.03.19.713075
Figure Lengend Snippet: Overview of the proposed pipeline. Each fundus image is preprocessed to produce a vessel-enhanced image and a morphological skeleton for persistent homology. EfficientNet-B3 provides 1536-d CNN features; TDA yields six topological descriptors concatenated to form augmented node representations. A topology-aware population graph connects similar images; two-layer GraphSAGE refines representations through neighbourhood aggregation for five-class DR grading.
Article Snippet: Feature extraction with EfficientNet-B3 scales linearly as O ( N ); for N = 88,702 (Kaggle DR), this requires ≈7 min on a single A100 GPU.
Techniques: