Parameter-Efficient Diabetic Retinopathy Grading with LoRA-Based Fine-Tuning of Vision Foundational Models

Diabetic Retinopathy (DR) is a serious eye complication that requires timely diagnosis to prevent vision loss. Conventional diagnostic methods are time-consuming and subjective. Artificial intelligence (AI) has emerged as a promising solution for accurate and early diagnosis, but existing models are large, expensive, and resource-intensive, limiting their practical implementation. This work uses Low-Rank Adaptation (LoRA), a parameter-efficient fine-tuning technique, to adapt four state-of-the-art vision foundation models, DINOv2, Swin, Deit, and ViT for DR grading. Our approach not only performs on par with the most specialized models but is also resource-efficient, possessing significantly fewer parameters. This enables deployment on resource-constrained devices and promotes wider clinical adoption. Additionally, we employed ensemble techniques with our fine-tuned models, further enhancing performance, with our best model achieving an accuracy of 83.70%. This work represents a significant step toward more accessible and reliable AI-based DR diagnosis, paving the way for broader implementation in clinical settings.