Age Prediction from Fundus Images by a Multi-Branch Swin-Transformer Model

Fundus images are photographs of the retina, which contains information related to the degree of aging in the human body. If the age predicted from a fundus image is higher than the actual age, there is a risk of disease, while a predicted age close to the actual age indicates relatively good health. Therefore, age prediction from fundus images plays an important role in disease prevention, diagnosis, and prognosis. However, it is still unclear which retinal tissues contain the most age-related information, presenting a challenge in designing deep learning models. This study proposes a deep learning model called the Multi-Branch Swin-Transformer, which uses attention maps to visualize the regions of focus, exploring which parts of the retina are influential in age prediction. On the AI Content dataset, an Root Mean Square Error (RMSE) of 4.21 was achieved. The study concluded that age prediction from fundus images is primarily influenced by four regions: the macula, vessels, Optic Disc (OD), and Optic Cup (OC). And the ranking of these regions in terms of influence on age prediction is macula, vessels, OC, and OD.