<![CDATA[Diabetic retinopathy (DR) is a leading cause of preventable blindness, where early and accurate detection is critical for effective intervention. While deep learning models have demonstrated strong performance in DR classification, their limited interpretability and inconsistent evaluation practices hinder clinical trust and deployment. This study proposes an explainable deep learning framework for DR detection based on MobileNetV2, complemented by Integrated Gradients for feature attribution. A curated dataset of 4,464 retinal images was constructed from publicly available sources through systematic preprocessing, including quality filtering, deduplication, and class balancing across five DR stages. To ensure robust evaluation, a multi-level validation strategy was employed, incorporating stratified train–validation–test splits and k-fold cross-validation. The proposed framework achieved 87.0% accuracy and an F1-score of 0.868, outperforming baseline models including EfficientNet-B0, DenseNet121, and VGG16. Beyond predictive performance, explainability was quantitatively evaluated using deletion and insertion metrics, demonstrating that Integrated Gradients provides more faithful feature attribution compared to Grad-CAM and LIME. Error analysis further reveals that misclassifications are concentrated between adjacent DR stages, reflecting the inherent difficulty of fine-grained disease progression modelling. The findings highlight that combining rigorous validation with quantitative explainability evaluation can improve the reliability and transparency of deep learning models for medical imaging. While results are promising, the framework is validated on publicly available datasets and requires further external clinical validation before real-world deployment.]]>
Copyrights © 2026
