AGRF-Net: anatomy-guided reliability fusion for multi-view OCT/OCTA retinal disease classification

Abstract

Optical coherence tomography (OCT) and optical coherence tomography angiography (OCTA) provide complementary structural and microvascular information for retinal disease assessment. However, subject-level fusion is challenging because disease evidence may be distributed across retinal slabs and the diagnostic relevance of individual projections varies across subjects. We propose AGRF-Net, a multi-view OCT/OCTA classification model that explicitly organizes projection-level features according to retinal anatomical correspondence and learns sample-specific view-contribution weights for structural–vascular evidence fusion. Instead of treating OCT/OCTA projections as exchangeable feature tokens, AGRF-Net preserves their modality identity and retinal slab order during cross-view relation modeling and adaptive view fusion. Specifically, the local–global representation module preserves projection-specific pathological cues while producing relation-oriented semantic features for cross-view interaction. Building on these representations, the anatomy-guided relation graph combines subject-specific feature similarity with a slab-correspondence-based anatomical prior refined during training, thereby constraining information propagation among corresponding and complementary OCT/OCTA projections. Finally, contribution-weighted aggregation integrates the relation-enhanced view embeddings into a subject-level representation. Experiments on OCTA-500 show that AGRF-Net achieves balanced performance in complete six-view OCT/OCTA classification. These results support explicit modeling of slab-level OCT/OCTA correspondence and adaptive view weighting for integrating complementary structural and vascular evidence.