Myriaxon Medical ResearchCataract detectionVersion by Audri

Cataract detection, Audri’s version

Audri’s version separates three outcomes: a normal eye, an immature cataract and a mature cataract. Choose an eye photo, press Run prediction, and the model marks what it finds. It runs on your device: the photo is never uploaded.

1 Choose an eye photo

or drop one on the picture, or paste it

No photo handy? Press a sample. The tag shows what the photo really is, so you can judge the model against it.

2 Patient details (optional)

These stay in your browser. They are only printed on the report image you can download.

3 Prediction

Loading the model... 0%

About this version.

The model
A YOLOv10-nano object detector, about 9 MB, that looks for the eye and sorts it as immature, mature, normal. The original application reports a mAP of 0.994 on its own test data. We have not reproduced that figure. The model in this demo is the author’s trained network converted to run in the browser, and gives the same answers as the author’s network run directly on the photo. (The original Gradio app fed the network its colours in the wrong order, which made it miss many cataracts; the apps are fixed, and this demo was never affected.)
What it can get wrong
It was trained on close-up eye photographs of one kind, and works best on those. On ordinary photos, clinic scenes or unusual lighting it may say “No detection”, and on any photo it can call a healthy eye a cataract or miss one. Confidence is how sure the model is, not the chance that the person has the disease.

A research demonstration, not a medical device. It does not diagnose anything and must not be used to decide on care. If you are worried about your eyes or your vision, see an eye doctor.

The sample photos are public photographs from Wikimedia Commons, some of healthy eyes and some of cataracts (the tag on each says which). They are not from the thesis data set, and a model trained on one kind of photo does not always recognise a cataract in another: try the dense white cataract to see a miss. Credits:

    Also see Habib’s version, built independently, or all Myriaxon Medical Research.