Abstract
BACKGROUND: Automated analysis of color fundus photographs can support scalable screening for diabetic retinopathy (DR), age-related macular degeneration (AMD), and glaucoma, but single-run reporting and accuracy-only summaries can mask clinically relevant instabilities and failure modes. METHODS: Using the public FIVES dataset, we benchmarked six deep learning configurations for four-class fundus screening (AMD, DR, glaucoma, normal): three DeepLabv3-backbone hybrids (ResNet50, DenseNet121, EfficientNet-B0) and three backbone-only classifiers. All experiments were evaluated using five independent stratified splits generated with different random seeds (n = 5 runs), each defining a distinct 20% held-out test set. Models were trained on the remaining 80% (training/validation), and all reported metrics are computed on the 20% test set of each run and summarized as mean ± SD across runs. Performance was summarized with accuracy, sensitivity, specificity, and one-vs.-rest AUC; we further characterized clinical behavior via row-normalized confusion matrices, per-class precision/recall/F1, and a screening-style binary triage setting (referable = AMD ∪ DR ∪ glaucoma vs. normal). RESULTS: Hybrid models consistently achieved higher discrimination than simple classifiers (AUC 0.969-0.979 vs. 0.908-0.920), despite similar accuracies (0.924-0.941). The selected model, DeepLabv3-DenseNet121, reached the highest AUC (0.979 ± 0.009). Class-wise analysis revealed strong performance for Normal (F1 0.970 ± 0.014) and Glaucoma (F1 0.896 ± 0.048), while DR was the main bottleneck (sensitivity 0.738 ± 0.117), with most DR errors redistributed to AMD (13.6%) and Glaucoma (12.0%) and minimal confusion with Normal (0.5%). In binary triage, the model achieved sensitivity 0.993 ± 0.011 and specificity 0.963 ± 0.034, with PPV 0.987 ± 0.013 and NPV 0.980 ± 0.032, and a stable referral rate (∼0.73-0.77) across runs. CONCLUSION: DeepLabv3-based hybrids provide a robust advantage in AUC for multiclass fundus screening on FIVES. The residual risk concentrates in the DR-AMD-Glaucoma decision boundary, suggesting that deployment-oriented policies should prioritize conservative handling of DR-adjacent cases while leveraging the stability of Normal predictions for screening workflows.