A deep learning system for detecting diabetic retinopathy across the disease spectrum

AI Generated Summary*

DeepDR, a multi-task deep learning system trained with transfer learning, rates fundus image quality in real time, flags and outlines retinal lesions, and assigns grades from mild non-proliferative to proliferative disease. Training used 466,247 images from 121,342 people with diabetes screened in Shanghai. Local testing gave an AUC of 0.934 for image quality across 200,136 images, 0.901 to 0.967 for four lesion types in a 4621-image subset, and 0.943 to 0.972 for the four grades among 178,907 gradable images. External grading AUCs spanned 0.916 to 0.970 across 209,322 images. Real-time feedback for 1294 older adults dropped the low-quality rate from 28.7% to 8.2% and raised mild-grade AUC from 0.880 to 0.933; lesion overlays also aided community graders.

*This summary was generated by AI and is published unedited. Oku does not alter these summaries. It may contain errors or omissions and is provided for general informational purposes only. Oku does not guarantee its accuracy, completeness, or reliability. For authoritative information, please refer to the original, peer-reviewed article.