Optimized Deep Learning Framework for Diabetic Retinopathy Based on Systematic Multi‐Stage Transfer Learning and Predictive Risk Stratification

ABSTRACT Diabetic retinopathy (DR) is a leading cause of preventable vision loss, highlighting the need for reliable and scalable diagnostic and risk assessment methodologies. This study presents an optimized deep learning framework for predictive risk stratification in DR through a systematic tuning and evaluation of six pre‐trained convolutional neural network (CNN) architectures. A combined multi‐source dataset of retinal fundus images was employed to support model development and generalization. Preprocessing techniques, including CLAHE‐based contrast enhancement and data augmentation, were applied to improve lesion visibility and reduce inter‐dataset variability. A three‐stage progressive fine‐tuning strategy was developed to adapt ImageNet‐initialized models to retinal pathology. This strategy enables efficient feature reuse and stable convergence across all evaluated architectures. Model performance was evaluated using Accuracy, Precision, Recall, and F1‐score, with EfficientNetB7 achieving the highest validation accuracy of 96% and consequently selected for risk prediction. To move beyond categorical classification, a survival‐inspired encoding of DR severity was introduced to estimate a continuous risk score aligned with disease severity, enabling stratified risk assessment rather than discrete class labels. This design extends conventional DR classification toward clinically meaningful risk‐based assessment. The results demonstrate strong diagnostic performance and clear risk stratification capability, underscoring the effectiveness of multi‐stage transfer learning for fundus‐based DR risk prediction.