Machine Learning Approaches for Detecting and Classifying Diabetic Retinopathy – A Survey

Abstract

Abstract— Diabetic retinopathy (DR), a common complication of diabetes affecting the small blood vessels in the retina, is a primary reason for vision impairment. Timely identification is essential in averting sight loss; however, the manual examination of retinal fundus images is labor-intensive and susceptible to inaccuracies. Machine learning (ML) offers a promising solution, with various algorithms demonstrating high accuracy in DR detection and classification. This paper surveys 15 recent studies on ML-based techniques for DR analysis. We compare and contrast the performance of different algorithms, including convolutional neural networks (CNNs), deep learning models like Swin Transformers, hybrid approaches combining CNNs with other methods, and feature fusion techniques. The reviewed studies utilize publicly available datasets and evaluate their model’s using metrics like accuracy, sensitivity, specificity, and F1 score. Our analysis reveals that CNNs, particularly ResNet50 architectures, achieve high performance in DR detection and classification. Deep learning models like Swin Transformers demonstrate even greater accuracy, while hybrid approaches and feature fusion techniques offer additional advantages in classification and computational efficiency. The results underscore the capacity of machine learning to automate diabetic retinopathy screening, offering the promise of enhanced accessibility and precision when contrasted with conventional approaches. Future research directions include exploring explainable AI techniques for improved model interpretability and developing robust models that can generalize well across diverse datasets.