Great collaboration with Paul Jäger and Luciana Ferrer.

Funding: @helmholtzimaging.bsky.social and HIDSS4Health

Similarly, the popular temperature scaling method is insufficient to compensate for prevalence shifts as it is accuracy-preserving.
4/

Using argmax blindly to convert softmax output to class scores can have devastating consequences.
3/

The key insights and recommendations are summarized here.

Importantly:
2/

Really excited about sharing this paper hot off the press: doi.org/10.1016/j.me...
Invited to MedIA as one of the top MICCAI 2023 papers it presents large-scale analyses and concrete solutions related to prevalence shifts in medical image classification: 1/

The value of the framework was demonstrated on a set of diverse use cases.

Congrats to Tim Adler, Jan Nölke and the whole team!
The work was performed in the scope of @helmholtzimaging.bsky.social and @ercresearch.bsky.social grant NeuralSpicing.
4/n

... propose a framework for application-focused validation of posteriors. 3/n

We show that many real-world applications involving inverse problems require a "mode-centric" perspective on the problem and 2/n

🗞️Hot off the press🗞️: doi.org/10.1016/j.me... published in MedIA. Many image analysis problems are ambiguous, but validation of solutions to inverse problems is far from trivial. In this paper we take inspiration from the field of object detection: 1/n

Hello Bluesky world🙋🏼‍♀️,

Let me start with good news: Superstar Shekoofeh Azizi, research lead at Google DeepMind, will visit us in Heidelberg next week and give a talk in our @ellis.eu / hd.ai / NCT data science seminar. You can register here: www.dkfz.de/en/datascien... (talk will not be recorded)