Lesion-Fidelity-Aware Synthetic Data Augmentation for Diabetic Retinopathy Grading Using Conditional GANs

Assessing the severity of diabetic retinopathy using retinal fundus photographs constitutes an ordinal classification task in which clinically significant lesions influence referral decisions. In small and imbalanced datasets, deep learning models frequently overfit the most prevalent grades, motivating the use of synthetic augmentation via generative adversarial models. However, photorealism does not guarantee lesion accuracy, as conditional generators may omit, invent, or misplace lesions, introducing label noise and reducing ordinal consistency. This study proposes a lesion fidelity–aware augmentation framework that filters generated samples using a lesion fidelity score based on segmentation consistency. A multi-lesion U-shaped convolutional neural network is trained on pixel-level lesion masks from the Indian Diabetic Retinopathy Image Dataset and used as a fixed evaluator. For each synthetic image generated by a mask-conditioned image-to-image translation model, lesions are re-segmented and compared with the conditioning mask using Dice agreement. The fidelity–utility hypothesis is evaluated under controlled conditions on a subset of ninety-eight images with fixed splits and a matched synthetic budget of fifty-six samples. High-fidelity samples improve ordinal agreement, whereas low-fidelity samples yield stronger class-balance metrics, indicating a trade-off between fidelity and robustness. To account for variability due to the small test set, a multi-seed evaluation with three seeds is performed, showing consistent trends across runs. These findings should be interpreted as a controlled proof-of-concept under limited data conditions.