Optimized multi-scale dual-modality feature pyramid network for early eye disease classification using fundus and OCT imaging

Early and accurate detection of eye diseases, including diabetic retinopathy, macular edema, age-related macular degeneration (AMD), and glaucoma, remains a major challenge in ophthalmology due to the inherent limitations of single-modality imaging. Fundus photography captures surface-level vascular and optic disc information but lacks depth, while optical coherence tomography (OCT) provides high-resolution cross-sectional details of retinal layers but misses surface-level cues. To overcome these shortcomings, we propose an optimized hierarchical dual-modality feature pyramid network (OHDM-FPN) that integrates fundus and OCT images for robust multimodal eye disease detection. The framework employs EfficientNetV2-Small as the encoder for fundus images and GhostNet with dilated residual blocks for OCT feature extraction. Each modality is processed through a feature pyramid network (FPN) with atrous spatial pyramid pooling (ASPP) to improve multi-scale representations. A cross-modality multi-scale fusion block (CM-MSFB), enriched with channel-spatial attention (CBAM) and deformable convolutions, aligns and fuses modality-specific features. Finally, a lightweight classification head predicts disease categories, while hyperparameter tuning via the Salmon Salar optimization (SSO) algorithm addresses class imbalance and improves generalization. Experiments were executed on a combined fundus–OCT benchmark dataset (EyePACS + Duke OCT). The proposed OHDM-FPN achieved 98.2% accuracy, 97.9% sensitivity, 98.7% specificity, 0.993 AUC, and 98.1% F1-score, outperforming state-of-the-art single-modality models by an average margin of 6.3% in accuracy and 5.8% in AUC. Moreover, the fusion mechanism improved early-stage disease detection rates by 9.1% compared to OCT-only baselines. Attention heat maps confirmed that OHDM-FPN focused on clinically relevant retinal regions, improving interpretability for ophthalmologists. These results demonstrate that OHDM-FPN provides a powerful, optimized, and interpretable multimodal solution for eye disease detection, offering significant improvements in diagnostic performance and early intervention potential.