IDEA Lab @ MIDL 2025

Symbolic picture for the article. The link opens the image in a large view.

July 2025 — Salt Lake City, Utah

 

A team from the IDEA Lab made a strong showing at the Medical Imaging with Deep Learning (MIDL) 2025 conference, presenting their most recent research on generative models.

Among the highlights:

  • “Generate to Ground” introduced a novel use of text-to-image diffusion models for grounding medical phrases in images. The method outperformed state-of-the-art discriminative approaches, doubling accuracy in key benchmarks.

    Authors: Felix Nützel, Mischa Dombrowski, Bernhard Kainz

  • “Can Diffusion Models Generalize?” tackled the pressing issue of privacy and fairness in medical generative models, proposing a new metric, t′, to assess generalization trade-offs. This work offers critical insights for secure and ethical synthetic data sharing.

    Authors: Mischa Dombrowski, Bernhard Kainz

  • “CRG Score” presented a clinically-informed metric to evaluate long-form radiology reports. By addressing class imbalance and boosting interpretability, it sets a new standard for textual evaluation in medical AI.

    Authors: Ibrahim Hamamci, Sezgin Er, Suprosanna Shit, Hadrien Reynaud, Bernhard Kainz, Bjoern Menze