Concepts

From a fundus photo to a cup-to-disc ratio

A fundus photograph is a colour picture of the back of the eye. GlaucomAI's structure model works in two stages, each a U-Net, a network designed for medical image segmentation.[11] Stage A finds the optic disc in the whole photo. The photo is then cropped to 2.2 disc diameters around it, and stage B outlines the disc and the cup on the crop. The vertical CDR is measured from those outlines.

A plausibility check refuses photos in which no believable disc is found, so the model reports “ungradable” instead of guessing.

PAPILA RET014OD, full fundus photo
1. Find the disc
Stage A scans the whole photo
Optic disc crop
2. Crop around it
2.2 disc diameters, 256 × 256
3. Outline disc and cup
Stage B predicts two masks
4. Measure
vertical CDR = 0.74
Open-data imagePAPILA RET014OD (open data, GPL-3.0), processed by GlaucomAI's two stages.
disc crop256 × 256 pxdisc + cup masks32326464128128256256512 channels, 16 × 16skip connections keep fine detailencoder: what is in the imageconvolutions + downsamplingdecoder: where exactlyupsampling back to 256 × 256
SchematicThe U-Net in both stages: the encoder shrinks the image to learn what is in it, the decoder expands it back to say where, and skip connections carry fine detail across.
Key facts
  • Trained on Chákṣu, 1,345 photos from three cameras, each outlined by five experts; this project used its 1,009 training photos.[12]
  • Tested on two other sources: PAPILA and HYGD.[13, 14]
  • Outline overlap is measured with the Dice coefficient, where 1.0 is a perfect match.[15]
In GlaucomAI

Disc overlap (Dice) 0.98 on Chákṣu's test photos and 0.93 on PAPILA. Glaucoma AUC 0.85 on PAPILA and 0.72 on HYGD, cameras the model never saw.

Open the live demo

References on this page

  1. 11.Ronneberger O, Fischer P, Brox T. U-Net: convolutional networks for biomedical image segmentation. MICCAI 2015, LNCS 9351:234–241. arXiv ↗
  2. 12.Kumar JRH, Seelamantula CS, Gagan JH, et al. Chákṣu: a glaucoma specific fundus image database. Sci Data. 2023;10:70. DOI ↗
  3. 13.Kovalyk O, Morales-Sánchez J, Verdú-Monedero R, et al. PAPILA: dataset with fundus images and clinical data of both eyes of the same patient for glaucoma assessment. Sci Data. 2022;9:291. DOI ↗
  4. 14.Abramovich O, Pizem H, Fhima J, et al. Hillel Yaffe Glaucoma Dataset (HYGD): a gold-standard annotated fundus dataset for glaucoma detection. PhysioNet, version 1.1.0. DOI ↗
  5. 15.Dice LR. Measures of the amount of ecologic association between species. Ecology. 1945;26(3):297–302. Google Scholar ↗
Previous
Optic disc and CDR
Next
OCT imaging