A trustworthy cross-domain AI framework for fundus disease classification using hybrid CNN fusion and supervised domain adaptation

Abstract

Deep learning–based systems for fundus disease classification often achieve impressive accuracy on internal datasets, yet their performance degrades markedly when applied to real-world clinical data. This limitation is primarily caused by domain shift arising from variations in imaging devices, illumination conditions, and population characteristics, which remains a key barrier to integrating these technologies into reliable clinical decision-support systems. To address this challenge, we propose FusionEye-Net, a hybrid deep learning framework that performs feature-level fusion of EfficientNet-B3 and ResNet-50 for four-class fundus image classification, namely cataract, diabetic retinopathy, glaucoma, and normal retina. A dedicated preprocessing pipeline—including circular fundus cropping, illumination normalization, contrast-limited adaptive histogram equalization (CLAHE), and adaptive gamma correction—was employed to reduce inter-device variability and standardize retinal appearance. FusionEye-Net was first trained on a curated internal dataset of approximately 4000 images, achieving an internal test accuracy of 99.24%. However, evaluation on an independent external dataset comprising 4640 images revealed a significant performance drop to 72.05%, highlighting the severity of cross-domain variability. To mitigate this degradation, a targeted supervised domain adaptation strategy was applied using a balanced subset of 2000 external images (500 per class). This adaptation improved external performance to 91.27% accuracy with a macro F1-score of 0.9139, reducing misclassifications by more than two-thirds. Model predictions and visual explanations were reviewed by a board-certified retina specialist to ensure clinical plausibility. Explainability analysis using Gradient-weighted Class Activation Mapping (Grad-CAM) and region-of-interest (ROI) contour mapping demonstrated that the adapted model consistently focuses on medically relevant structures, such as lens opacity in cataract, microaneurysms in diabetic retinopathy, and optic disc cupping in glaucoma. In summary, the results demonstrate that FusionEye-Net combines strong internal performance with substantially improved cross-domain generalization through lightweight adaptation, underscoring the critical role of external validation and domain-aware fine-tuning in formulating trustworthy, adaptive AI-assisted decision-support tools that can safely augment automated ophthalmic screening workflows.