Research · 2022 · Research at KJ Somaiya College of Engineering · TechRxiv preprint
Exudate Segmentation in Retinal Fundus Images
Where exactly is the eye leaking?
A Recurrent Residual U-Net that segments exudates, an early sign of diabetic retinopathy, with 93.20% sensitivity and 99.80% specificity at the lesion level on E-Ophtha.
Problem
Exudates are among the earliest visible signs of diabetic retinopathy and are small enough to miss on a fundus photograph.
Approach
- A Recurrent Residual U-Net: stacked recurrent units in each layer and residual skip connections between layers, so a deeper network still trains.
- Trained on IDRiD; tested on IDRiD, E-Ophtha and DiaretDB1.
- Metrics computed at pixel, exudate and image level.
Stack
- Python
- TensorFlow
- Keras
- OpenCV
Outcome
- 93.20% sensitivity and 99.80% specificity on E-Ophtha at the exudate level