Automated Diabetic Retinopathy Severity Grading using Self-Supervised Vision Transformers: A Comprehensive Study on Transfer Learning for Medical Image Classification

Diabetic retinopathy (DR) causes about 93 million people in the world to lose their sight even though the condition can be treated. Proper medical treatment requires doctors to assess patient severity yet ophthalmologists need to conduct manual tests which result in high costs and extended working hours because the results vary among different doctors. The study presents a DR severity grading system which operates automatically by using DINOv2 to process APTOS 2019 Blindness Detection dataset after training on 142 million unlabeled images through self-supervised vision transformer technology. The proposed pipeline applies Ben Graham normalization to improve retinal contrast through attention-based feature pooling and focal loss with class weighting which helps solve the severe class imbalance issue that exists between the five DR severity grades. The evaluation process uses two model variants which include an efficient model (DINOv2-Base, 384×384) that achieves 87.49% accuracy with a quadratic weighted kappa of 0.734 and an accurate model (DINOv2-Large, 518×518) that achieves 90.62% accuracy with a QWK of 0.762. The combination of mixed-precision training with gradient accumulation provides a workable solution for deployment purposes. The findings show that self-supervised visual representations can be effectively utilized in medical imaging which enables telemedicine systems to use diabetic retinopathy screening programs.