Automatic glaucoma screening from a fundus photograph

Deep learning finds the optic disc, outlines disc, cup and parapapillary atrophy, measures the cup-to-disc ratio and the neuroretinal rim, and returns a calibrated glaucoma probability with a heat map and an image-quality check. The models run entirely in your browser; your photo is never uploaded.

0.978 / 0.904disc / cup Dice, 522 test photos
0.031vertical CDR error (classical: 0.178)
0.877AUC on clinical diagnoses (classical: 0.602)
92.7%vs best of 6 ophthalmologists 91.1%

Try it

Pick a held-out test photo (never used for training) or drop your own fundus photograph. The first run downloads the models (about 190 MB, cached afterwards).
1. Find optic disc
2. Disc & cup
3. Rim & PPA
4. CNN + heat map
5. Decision & quality

How it works

1. LocaliseU-Net (ResNet-34) marks the optic disc on the whole photo.
2. SegmentU-Net outlines disc and cup on a 3× crop; a U-Net++ refines the cup on a tight 1.5× crop.
3. MeasureVertical, horizontal and area cup-to-disc ratio; rim width in 360 directions (ISNT).
4. AtrophyPPA segmentation trained on PALM, plus a ring CNN that reads the band around the disc.
5. DecideResNet-50 on the disc region fused with CDR and rim by a calibrated logistic model.
6. CheckHeat map of the decision and an uncertainty-based image-quality warning.

Key results

Classical vs deep learning
Classical image-processing pipeline vs this system on the same 522 test photos.
ROC curves
Glaucoma detection on 411 clinically labelled test eyes.
Comparison with ophthalmologists
Accuracy on the base paper's task compared with each of six ophthalmologists.
Robustness
Behaviour under blur, illumination change, compression, low resolution and noise.