Machine learning models have proven their use in the medical field, assisting physicians in early diagnosis of serious diseases. Diabetic retinopathy is one of such diseases that could benefit from early
detection, but the amount of data available to train accurate models is limited. This challenge could be tackled through the use of syn thetic data in the training of machine learning models. This paper
proposes a novel procedure for the generation of synthetic data-sets for diabetic retinopathy that accurately reflect the characteristics and complexities of real-world medical data-sets. By employing a
systematic literature review in combination with an expert study, this research aims to improve the data generation that underpins the performance and reliability of AI systems for diabetic retinopa thy detection, enhancing early detection and treatment accessibility. Special attention is paid to inherent, human-centered issues in the synthetic data generation to mitigate representation risks.
