LVD Dataset containes retinal images?

[pnderitu]

I am a UK based researcher in ophthalmology curious to know if any ophthalmology datasets were used for LVD datasets, specifically images of the back of the eye (retina) as some are open access. If so approximate numbers.

This would be very valuable information since eye research are finetuning the model for eye related tasks and the models perform well but it could be contaminated by training data.

Thank you for your insights.

[pnderitu]

@patricklabatut

[patricklabatut]

We have not explicitly put retinal images in the training dataset LVD-142M. Whether such images actually exist in this dataset is hard to answer without having the corresponding datasets of these evaluation tasks and looking explicitly for them in LVD-142M.

[pnderitu]

Thank you very much for taking the time to answer my question, @patricklabatut !

FYI - The eye datasets in question are:
Kaggle EyePACS: https://www.kaggle.com/competitions/diabetic-retinopathy-detection/data
Messidor: https://www.adcis.net/en/third-party/messidor2/
IDRID: https://idrid.grand-challenge.org/

A follow-up question is whether there are any retinal images in LVD-142M (perhaps a nearest neighbour search using some example retinal images would reveal this).

It would be of great interest (to me as an ophthalmologist at least) to know as many papers (in ophthalmology) report good linear probing performance on eye disease with DinoN series, but don't think we have explicitly assessed if retinal images are in the training dataset.