AI systems in medicine make decisions with far-reaching consequences. It is important to understand how they go about doing this. In a new study, Farhad Nooralahzdeh (ZHAW) and researchers from the University of Zurich, the University Hospital of Zurich, ETH Zurich, Stanford University and Kobe University have investigated how medical image-text models can be specifically influenced: certain internal features can consistently improve model performance, whilst others help to suppress undesirable or misleading behaviour.
The result is greater control, improved traceability and more robust AI systems in clinical settings. The work was carried out as part of the “3D Vision-Language Model for Radiology” project, funded by the Swiss AI Initiative and Radical, which is receiving funding in the 4th Project Call.