A Survey On Publicly Available Fundus Image Datasets For Diabetic Retinopathy Screening

Diabetic Retinopathy (DR) which is the primary cause of visual impairment around the world, largely depends on fundus images for diagnosis and severity evaluation. The existence of diverse and high-quality fundus image datasets is critical in determining the progression of research and evaluating the ability of newly developed machine learning models for DR detection and classification. This survey goes through publicly available fundus image datasets and provides great insights to new researchers in this field. The study conducted a thorough search of the Scopus online library and the academic databases, encompassing papers from 2019 to 2024. After rigorous evaluation of the abstracts, methodologies, experiments, and conclusions of 60 papers, 25 datasets were identified from 60 papers. These datasets are both useful in understanding the progression of DR and as training and validation tools for machine learning algorithms. Because of the distribution shift, the study also characterizes key challenges of the problem including annotation precision and the generalizability of models across datasets. It also highlights the potential of ultra-widefield imaging for the detection of peripheral retinal abnormality and of the multi-modal approach which involves advanced imaging techniques with fundus photography. These approaches identify promising directions in enhancing the current DR detection, grading, and segmentation methods.