Multiclass Diabetic Retinopathy Classification based on Low-Rank Adaptation

Abstract

Diabetic retinopathy is a vision impairment associated with diabetes mellitus, a common medical condition. Retinographic imaging analysis of retinal fundus images is the predominant method of diagnosing this eye complication. Recent advances in deep learning show effectiveness in detecting diabetic retinopathy, rivaling the diagnostic accuracy of human inspection. However, the success of these deep learning methods largely depends on the architecture model. In our research, we introduce a method for training a deep learning model using a Low-Rank Adaptation (LoRA) technique to differentiate between three stages of diabetic retinopathy: no diabetic retinopathy, non-proliferative diabetic retinopathy, and proliferative diabetic retinopathy. LoRA significantly reduces the training parameters needed by employing a low-ank representation. Our experimental results on four data sets of different sizes show that the LoRA model delivers top-tier results in classifying stages of diabetic retinopathy.