And another one: We show how to use a diffusion model for #fundus images for #counterfactual reasoning in #ophthalmology! If you ever wondered how the fundus images of your patient had looked, if he had been suffering from diabetic retinopathy, check this out ⬇️ journals.plos.org/digitalhealt...

In our model, we use the gradients of a robust disease classifier and regularization towards the original image to guide the diffusion process

This allows to generate high quality counterfactual images for individual patients, e.g. answering the question how an image had looked if the patient had been healthy or if she had been sick. The induced lesions look highly realistic, and the vessel structure is kept intact.

Importantly, we do not simply compute quality metrics, but perform a highly sensitive three-way odd-one-out study to assess the realism - even for ophthalmologists, it is hard to figure out, which images have been generated by our algorithm!

Joint work with PhD students Indu Ilanchezian & Valentin Boreiko, as well as Ziwei Huang @msayhan.bsky.social @lisakoch.bsky.social Matthias Hein and Laura Kühlewein!