Managing Diabetic Retinopathy with Deep Learning: A Data Centric Overview

A research team from multiple institutions, led by Shramana Dey, has published a comprehensive review paper titled 'Managing Diabetic Retinopathy with Deep Learning: A Data Centric Overview' on arXiv. The paper addresses a critical bottleneck in medical AI: the limited availability of high-quality, standardized datasets for training deep learning models to detect and grade diabetic retinopathy (DR), a leading cause of vision loss. The researchers conducted a comparative analysis of existing fundus image repositories, evaluating their usability for key clinical tasks like binary classification (disease/no disease), severity grading, lesion localization, and multi-disease screening.

The review systematically categorizes datasets by size, accessibility, and annotation type—such as image-level labels versus detailed lesion-level markings. It highlights persistent, significant gaps that hinder the development of clinically reliable AI, including geographically narrow data that lacks diversity, inconsistent annotation standards between datasets, and a severe shortage of longitudinal data tracking disease progression over time. As a case study, the paper examines a recently published dataset to illustrate broader challenges in curation. The authors conclude with concrete recommendations for future dataset development, emphasizing the need for standardized lesion annotations and more diverse, longitudinal data to build truly explainable and trustworthy AI screening tools.