Rwandan researchers explore AI to improve eye disease screening

Rwandan researchers are exploring how artificial intelligence (AI) could help health workers analyse eye images and identify signs of disease more quickly, with the long-term goal of making screening more accessible, particularly in areas with limited access to specialist care. The research focuses on identifying blood vessels in images of the retina, the light-sensitive tissue at the back of the eye. By training AI to recognise these vessels, the researchers hope to develop tools that could eventually support screening for diabetic retinopathy, an eye disease caused by diabetes that can lead to vision loss. ALSO READ: AI in healthcare: Why Rwanda should innovate and not just consume The team believes the technology could help address challenges faced by patients who have limited access to eye specialists and advanced medical equipment. If successfully developed and clinically validated, an AI-powered screening tool could eventually support health workers in clinics and health centres, including facilities far from major hospitals. “AI in medicine is something that is currently gaining a lot of attention around the world and also in Rwanda,” said Jean De Dieu Niyonteze, one of the researchers. He said the team is exploring how the technology could help address diabetes-related eye disease. “We are working on an AI project that addresses diabetes-related eye disease, particularly in the retina. This is called diabetic retinopathy,” he said. The research, however, is still at an early stage and has not yet produced a medical device ready for use in hospitals. What the research found The study, titled Experimental Evaluation of Public Retinal Vessel Segmentation Datasets (2020–2025) with Deep Learning: An Empirical Study, was presented at the 2026 International Conference on Advanced Research in Computing and published through the Institute of Electrical and Electronics Engineers (IEEE). It examined 11 publicly available datasets containing retinal images. The researchers used an AI model known as U-Net++ to test how accurately it could identify and separate blood vessels in the images. In simple terms, the team wanted to establish whether the AI could examine an image of the back of an eye and accurately trace the blood vessels visible in it. ALSO READ: New study explores how wearable sensors, AI could help save mothers and newborns Accurately identifying these vessels is an important step towards developing computer-based systems that can analyse retinal images for signs of eye disease. According to the researchers, the AI model recorded a Dice score of 0.80 and an overall accuracy of 0.97 when tested on the Retinal Vascular Tree Analysis (RETA) Benchmark dataset. The Dice score measures how closely the blood vessels identified by the AI match those previously marked in the images by researchers. A higher score indicates a closer match. The researchers also reported varying results across other datasets, highlighting the need for further testing before the technology can be considered for clinical use. They attributed the differences in performance to several factors, including image quality and size, the number of images available, and the accuracy of the original blood vessel markings used to train and evaluate the model. The findings highlight the importance of testing AI models on different datasets before they can be considered for use in healthcare settings. From research to clinical application Niyonteze, who studied Artificial Intelligence Engineering at Carnegie Mellon University Africa before pursuing Business Analytics at Emory University, said the findings represent an early step towards developing a tool that could analyse retinal images collected from patients in Rwanda. The team now plans to develop more advanced versions of the AI models and evaluate their performance using images they have not previously encountered. “We have already obtained our first results, which we presented at the International Conference on Advanced Research in Computing 2026,” he said. “The next stage is to develop what is called the final models and conduct inference evaluation.” Inference evaluation involves testing a trained AI system on new data to assess how well it performs beyond the images used during its development. ALSO READ: Can AI unlock Africa’s healthcare bottlenecks? For the researchers, a key question is whether an AI system trained on international datasets can accurately analyse retinal images from Rwandan patients. If the models perform well, the team hopes to develop a prototype that could eventually be tested in healthcare settings. Such a tool would need to undergo appropriate clinical validation and meet applicable regulatory requirements before being introduced for medical use. Beyond the technical performance of the models, the researchers are also considering how the technology could be made accessible and practical for health workers. Benny Uhoranishema, another member of the team who has a background in computer engineering and machine learning, said any eventual tool would need to be simple enough for doctors and other health professionals to use without requiring specialised knowledge of AI. “You understand that a doctor is not going to sit down and read the code or understand the technology behind it,” he said. “What matters is that the technology works in the same way as other technologies they already use, such as radiography and others.” He said the team could develop a dashboard that presents the AI's findings in a clear and accessible format, allowing doctors to interpret the results without needing to understand the technology behind them. Addressing gaps in specialist care The researchers emphasised that their current findings do not mean the AI can already diagnose diabetic retinopathy. The study focused specifically on identifying blood vessels in retinal images, rather than detecting or diagnosing the disease itself. Their broader project seeks to establish how this capability could eventually contribute to screening for diabetic retinopathy and support the work of healthcare professionals. The team is continuing to refine the models and plans to present further findings in February 2027. Niyonteze said the wider goal is to explore whether AI could help address gaps in access to specialised healthcare, particularly for patients who may struggle to access eye specialists. “We have a diabetes problem around the world, and Rwanda is no exception,” he said. “At the same time, there is a shortage of doctors and specialists. We believe this AI could help address that problem.”

News

Rwandan researchers explore AI to improve eye disease screening