A novel bilateral cross-attention network for multi-label fundus disease diagnosis

Multi-label fundus disease classification aims to assign multiple ocular disease labels to bilateral fundus images, enabling automated screening and clinical decision support. In contrast to most existing computer-aided systems that process each eye independently and rely on shallow feature fusion, we explicitly model binocular structure and adaptively combine information from both eyes to handle complex, real-world fundus scenes. To this end, we introduce DualCrossAttnNet, a multi-label fundus disease classification network specifically designed for binocular analysis. In particular, we obtain high-resolution bilateral representations using an EfficientNet-B2 backbone and a cross-attention module that jointly reasons about spatial and channel-wise interactions between the left and right eyes. We further employ a gated fusion mechanism to adaptively weight the contributions of each eye, and an SE attention module to recalibrate channel responses before global aggregation. Coupled with a fundus-oriented preprocessing pipeline and a GeM-based classifier, the proposed framework can accurately predict multiple co-occurring ocular diseases from complex clinical fundus images. On the public ODIR-2019 dataset, DualCrossAttnNet substantially improves multi-label classification performance, achieving 88.20% accuracy, 90.98% F1, and 98.49% AUC, with a composite score of 92.73%. These results surpass recent CNN, GNN, and Transformer-based baselines by up to 20.37 percentage points in composite score, demonstrating that DualCrossAttnNet is an effective and scalable solution for intelligent fundus disease diagnosis.