The development of deep learning technology and high-performance computers in recent years has enabled significant improvement in the accuracy of image recognition, and efforts are currently being made toward the development of ophthalmic artificial intelligence (AI). In this study, deep learning was applied to fundus image analysis. The fundus images were automatically classified by their severity index and the recognition accuracies for Japanese and American data were compared. We trained a deep convolutional neural network and a support vector machine using a data set consisting of 35,126 open-source American clinical fundus images, which were graded for diabetic retinopathy by licensed ophthalmologists. The grading accuracy of the trained AI model was evaluated by comparison with its performance for 200 Japanese fundus images obtained at Keio University Hospital. The trained AI model exhibited a sensitivity of 81.5% and specificity of 71.9% for the American validation data set, and a sensitivity of 90.8% and specificity of 80.0% for the Japanese data set. This indicates that the proposed AI program developed using American image data can be applied to fundus photographs of Japanese subjects and thus represents an interracial screening model. It can be used for screening tests and applied to telemedicine systems.
ASJC Scopus subject areas
- コンピュータ サイエンスの応用