In this paper, we propose a dual-modal multi-scale Siamese network, DMS-Net, for binocular fundus image classification. DMS-Net extracts deep semantic features from paired fundus images by utilizing a weight-sharing Siamese ResNet-152 backbone. To address issues such as lesion boundary ambiguity and scattered pathological distribution, we introduce a multi-scale context-aware module (MSCAM) that integrates adaptive pooling and attention mechanisms. In addition, we effectively combine global context and local edge features by enhancing cross-modal interactions using spatial-semantic recalibration and bidirectional attention via the dual-modal feature fusion (DMFF) module. When evaluated on the ODIR-5K dataset, DMS-Net achieves state-of-the-art performance with an accuracy of 82.9%, recall of 84.5%, and Cohen's kappa of 83.2%, demonstrating a superior ability to detect symmetric pathologies and advance clinical decision-making for ophthalmic diseases.
