Multi-Stage Diabetic Retinopathy Classification Using Deep Learning: Trends, Challenges, and Future Directions

Abstract

Diabetic Retinopathy (DR) is a primary cause of vision loss globally in patients suffering from diabetes. Timely detection of the disease can reduce the chances of blindness. Fundus images are studied by the ophthalmologists for any signs of the disease. Existing methods for detecting DR include manual study by the experts or usage of artificial intelligence-based models. But these methods encounter several issues, like inaccurate detection, time consumption, and cost, including the lack of stable datasets and consistent image processing techniques. Owing to the damage these lesions do to the retina, it becomes essential to look for an efficient and accurate model for diagnosing DR. Deep neural networks help us to segment the fundal images to focus on the retinal areas, helping us to look for early signs of the disease. If DR is detected at an earlier stage, the risk of complete blindness can be reduced. This paper presents a comprehensive survey of existing models for diabetic retinopathy detection, with an intention of identifying opportunities for improvement to develop a more robust and stable model that can efficiently and accurately perform the classification. It also includes a detailed discussion on various datasets used for detection/classification of the disease.