As Artificial intelligence is transforming healthcare, making it robust and reliable in real-world clinical settings remains a challenge. In this insightful DLD Future Hub session, Professors Katharina Breininger (University of Würzburg) and Reinhard Heckel (TU Munich) explore both challenges and opportunities of integrating AI into medical imaging and established healthcare systems.
In controlled studies, medical AI promises superhuman performance – yet consistently fails when deployed in real-world clinical settings, Reinhard Heckel notes.
The explanation lies in the complexity of different applications and environments. “It’s a really multifaceted, faceted problem and challenge”, Katharina Beininger says. “For example, in hospitals, we do not have just a single vendor that provides medical devices.” Similarly, there are no “standard patients” that would make things easier for AI systems that rely on average values, pattern recognition, and statistics.
The rise of foundation models offers hope, as they “alow us to have a model that starts already somehow intelligent”, Breininger notes. “So we don’t need to start training from scratch, but we already have, like, a 10-year-old child instead of a toddler that we’re trying to train.”
However, such models require vast, diverse data sets for reliable performance – and these are hard to obtain in medicine due to privacy and infrastructural constraints, both experts stress. “Access to medical data is a huge challenge”, Breininger says.
Watch the video to explore the progress of AI in healthcare further.
