LRRNet: A lightweight region-refined network for retinal disease classification from optical coherence tomography

With the wide clinical use of optical coherence tomography (OCT), lightweight and efficient deep learning models for retinal disease detection are in urgent demand. To address this challenge, we propose LRRNet (Lightweight Region-Refined Network), a lightweight multi-scale architecture for retinal OCT classification. Specifically, LRRNet is built upon a Selective Kernel (SK) backbone integrated with Dilated Context Fusion (DCF) blocks to capture multi-scale contextual representations without reducing spatial resolution. To suppress background noise and highlight pathological structures, a Multi-Band Adaptive Region-Aware Feature Refinement (MB-ARAFR) module is introduced, which performs vertically region-aware feature enhancement through overlapping band aggregation together with channel–spatial gating and a global attention branch. Furthermore, mid-level and deep features are pooled via Generalized Mean (GeM) pooling and concatenated to form a compact multi-scale embedding. The fused representation is then adaptively recalibrated by an Instance-Level Refinement Gate (IRG) prior to classification, enhancing feature discriminability at the single-instance level without introducing explicit inter-sample relational modeling. To optimize the network, we employ a two-stage coarse-to-fine training strategy incorporating mixup regularization, auxiliary supervision, and exponential moving average (EMA) to stabilize optimization and refine classification boundaries. Under standard single-view inference, LRRNet achieves an outstanding overall accuracy of 98.14% on the 8-class retinal OCT dataset (OCT‑C8), which outperforms existing state-of-the-art approaches while maintaining an extremely compact model size of merely 1.30 M parameters. These results demonstrate that LRRNet provides a highly accurate and computationally efficient solution for retinal OCT classification, supporting its potential for further deployment-oriented evaluation.