Enhancing early diabetic retinopathy detection: a fine-tuned NASNet framework with robust validation and external generalization

Abstract

Early detection of diabetic retinopathy (DR) is critical for preventing irreversible vision loss among patients with diabetes mellitus. Automated screening systems based on deep learning have demonstrated substantial promise in improving diagnostic efficiency and scalability. This study proposes a robust multiclass DR classification framework using a fine-tuned NASNetLarge architecture applied to retinal fundus images. The model leverages transfer learning from ImageNet and incorporates a structured fine-tuning protocol to adapt high-level features to DR-specific patterns. Experiments were conducted on the publicly available APTOS 2019 Blindness Detection dataset, comprising five DR severity classes. The model was evaluated using accuracy, precision, recall, F1-score, sensitivity, specificity, and receiver operating characteristic–area under the curve (AUC) metrics. To ensure robustness, experiments were repeated across five independent runs and reported as mean ± standard deviation. External validation was performed using the EyePACS dataset to assess generalisability. The fine-tuned NASNet model achieved 97.52% ± 0.42 accuracy with strong per-class AUC values (≥ 0.98). External validation yielded 93.42% accuracy and an AUC value of 0.941, confirming generalization capability. Comparative analysis against EfficientNet-B4, DenseNet121, and Vision Transformer (ViT-B16) demonstrated superior performance under identical conditions. The proposed framework shows strong potential as an automated DR screening tool for early-stage detection and large-scale population screening.