Retina Net-Health: Transformer – Based Retinal Image Classification for Diabetic Retinopathy Detection

Diabetic retinopathy (DR) is one of the leading reasons for preventable blindness, but the detection of the disease at the appropriate time is still hampered due to the lack of ophthalmic graders trained to diagnose DR as well as the variability in the quality of images used to perform screening. This work is aimed to overcome the problem of late and inconsistent diagnosis of DR and provides Retina Net-Health, a transformer-based retinal image classification framework to enhance automated detection accuracy and clinical reliability. The idea here is to develop a hybrid architecture that is capable of capturing subtle patterns of lesions, but also high sensitivity for referrable DR. The methodology includes CNN-based multi-scale feature stem and lesion-aware transformer tokenization, dynamic loss weighting and lesion-guided augmentation to make it more robust. Evaluated on EyePACS and an external clinical test set, the model produced an AUC of 0.982, sensitivity of 0.960 and F1-score of 0.920, which is better than five state-of-the-art baselines. These results demonstrate that lesion-guided token modeling and hybrid transformer reasoning achieve huge improvements in the diagnostic testing performance. Retina Net-Health is a clinically viable method of scalable DR screening, and is a pathway to discussing generalised interpretable retinal disease assessment.