Using Retinal Fundus Photographs
Professor Park Sang-joon at Seoul National University Bundang Hospital announced on April 14 that his research team has developed a deep learning algorithm accurately predicting ages and genders from retinal fundus photographs.
In developing the algorithm, the team used 412,026 retinal fundus photographs accumulated at the hospital. In recent tests, the algorithm succeeded in predicting general males’ and females’ ages with an average error of 3.1 and those of diabetes and hypertension patients with an average error of 3.6 or less. The accuracy of the algorithm is higher for up to 60, which is an age group showing more change in eyeball as the age increases, and the average error of the algorithm does not exceed 2.9 in any age group. When it comes to genders, the prediction accuracy is 96 percent or more regardless of underlying medical conditions.
The new algorithm is meaningful in that it retains its accuracy even in the presence of underlying medical conditions. This implies that a change in eyeball attributable to those conditions such as diabetes and hypertension and a change in eyeball attributable to aging are unique and distinguishable in relation to each other although they appear to be similar to each other.
Details of the research are available on the Mar. 12 online edition of the Scientific Reports journal.