Erasing and Refining Discriminative Features for CNV Subtype Classification in OCT Images

Choroidal neovascularization (CNV) subtype classification from optical coherence tomography (OCT) images is clinically important because treatment response and disease prognosis vary across subtypes. However, automated classification remains challenging owing to subtle inter-class differences and considerable imaging noise. We propose an erasing-refining discriminative feature network (ERDF-Net) that mitigates noisy dominant activations and reveals fine-grained structural cues. The model perturbs salient regions to facilitate subtle feature learning, restores clean salient information, and fuses both representations via channel-spatial attention to form a coherent and discriminative embedding of CNV morphology. Experiments on a clinical OCT dataset show that ERDF-Net consistently surpasses state-of-the-art fine-grained and erasing-based methods across multiple metrics. Ablation and visualization analyses further confirm the benefit and interpretability of controlled salient suppression and refined feature fusion. ERDF-Net provides an effective and reliable solution for fine-grained medical image classification.