Automatic detection of 39 fundus diseases and conditions in retinal photographs using deep neural networks
Problem
Detecting multiple retinal fundus diseases and conditions early enough to prevent irreversible visual impairment, previously addressed only one disease at a time.
Why it matters
Untreated fundus diseases cause irreversible blindness and specialist screening capacity is limited.
What the AI did
A deep learning platform classified 39 referable fundus diseases/conditions from 249,620 labeled images, then was tested on external multihospital data, public data, and a tele-reading application.
Results
Frequency-weighted average F1 0.923, sensitivity 0.978, specificity 0.996, AUC 0.9984 on primary test set; reached average level of retina specialists; high efficiency on external multihospital and public data tests.
Impact
Specialist-level multi-disease retinal screening could enable scalable tele-reading applications.
Record
| Venue | Nature Communications |
|---|---|
| Published | 2021 |
| Work type | article |
| Authors | Ling‐Ping Cen, Jie Ji, Jianwei Lin, Si-Tong Ju, Hong-Jie Lin, Tai-Ping Li, Yun Wang, Jian-Feng Yang et al. |
| Citations | 346 (OpenAlex, at import time) |
| References | 49 |
| Model or method | deep learning platform (DLP), deep neural networks |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1038/s41467-021-25138-w |
| Open access | gold |
| Keywords | deep learning, fundus photographs, retinal diseases, diagnostics, multi-label classification |
Importance rating
significant — thousands to millions affected
Untreated fundus diseases cause irreversible visual impairment; specialist-level multi-disease screening could enable scalable triage in remote areas worldwide.