The prevalence of diabetes is on the rise steadily around the globe. Diabetic retinopathy (DR) is a result of damage to the blood vessels in the retina due to diabetes and its fast treatment is crucial for preventing possible blindness. The diagnosis of DR is done mostly using a comprehensive eye exam, where the eye is dilated for better inspection. Analysis by an ophthalmologist is prone to human error and thus automatic and highly accurate detection of DR is preferred for an earlier and better diagnosis. It is important, however, that automatic detection be accurate for all data collected from patients of different geographic and ethnic backgrounds. In this paper, the automatic detection of DR with a deep learning algorithm is analyzed when geographic and ethnic information of the patients is also integrated into the architecture. It is shown that robust and generalizable DR detection performance is linearly related to the correlation of geographic and ethnic patient information between the training and the testing datasets. The deep learning model created eliminates geographic variation in the detection and works for patients of all ethnicities.
Deep learning, diabetic retinopathy, ethnicity, fundus images, geographic variation
SERENER, ALI and SERTE, SERTAN
"Geographic variation and ethnicity in diabetic retinopathy detection via deeplearning,"
Turkish Journal of Electrical Engineering and Computer Sciences: Vol. 28:
2, Article 6.
Available at: https://journals.tubitak.gov.tr/elektrik/vol28/iss2/6