Pre-training weight

[xingyuzhang1338]

Hi, thank you for your outstanding contributions, and I would like to ask what are the pre-training weights that you provide? Is it a pre-training weight for all data on eye disease?

[jusiro]

Hello @xingyuzhang1338,

I'm sorry for not getting back to you sooner.

FLAIR is pre-trained in an assembly dataset of color fundus images compiled from open-access data sources. In such datasets there are many different eye diseases represented, both specific diseases (e.g. glaucoma, diabetic retinopathy, or macular edema) and severity grades for specific ones (e.g. diabetic retinopathy grades). So yes, it is the same model weights for handling all the represented tasks.

You can find specific details in the manuscript. Concretely, Table 1 contains the employed datasets and appearing categories, whose abbreviations are specified in Appendix A. FLAIR is not pre-trained in all data from Table 1, since we left out a few datasets for testing. You can find more details in Section 4.2 of the main paper.

Kind regards.