Early diabetic retinopathy screening using a fine-tuned deep‑learning NASNet framework; a technical article detailing model development, validation, and generalization.
Key Takeaways
- Fine-tuned NASNetLarge model achieved 97.52% ± 0.42 accuracy on the APTOS 2019 dataset
- External validation on EyePACS produced 93.42% accuracy and an AUC of 0.941
- NASNetLarge outperformed EfficientNet-B4, DenseNet121, and ViT-B16 under identical evaluation conditions