Multi-categorical deep learning neural network to classify retinal images: A pilot study employing small database
importance 4/10
partially solved
Retinal Imaging and Analysis ophthalmology stage: pilot
Problem
Automated detection of multiple retinal diseases from fundus photographs using computer-aided diagnosis.
Why it matters
Retinal disease detection from fundus images could support clinical diagnosis.
What the AI did
A deep learning CNN (VGG-19-based random forest transfer learning, MatConvNet) classified fundus photographs from the STARE database into 10 categories.
Results
For all 10 categories: 30.5% accuracy, RCI of 0.052, Cohen's kappa of 0.224; performance diminished as category count increased
Impact
Pilot study demonstrating feasibility but low accuracy for multi-disease classification on a small database.
Record
| Venue | PLoS ONE |
|---|---|
| Published | 2017 |
| Work type | article |
| Authors | Joon Yul Choi, Tae Keun Yoo, Jeong Gi Seo, Jay Jiyong Kwak, Terry Taewoong Um, Tyler Hyungtaek Rim |
| Citations | 282 (OpenAlex, at import time) |
| References | 67 |
| Model or method | VGG-19 architecture with random forest transfer learning, implemented in MatConvNet |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1371/journal.pone.0187336 |
| Open access | gold |
| Keywords | deep learning, retinal diseases, fundus images, VGG-19, STARE database, classification |
Importance rating
4
real but narrower advance
Pilot on small STARE database; low multi-disease accuracy (30.5% for 10 categories) means feasibility demonstration only, limited current clinical value.