In the field of medical image analysis, the ultimate goal is to improve patient outcomes. Machine learning can help to achieve this goal by

• Accelerating and simplifying the analysis of medical images through (partial) automation of diagnosis, outcome prediction, image quantification, and image reconstruction.
• Developing technology which enables completely novel clinical workflows which are not possible without AI support.
• Extraction of new clinical knowledge from large image databases, which can inform future clinical decisions, treatments and drug trials.

Even though tremendous progress has been made for all of those points in research settings, surprisingly little of this technology has made it into medical practice. One reason for this is that the medical domain is an extremely high-stakes application field with extraordinary demands on robustness of algorithms. Another is that algorithmic outputs are not suitable for clinical decision-making if neither the patient nor the doctor can understand the reasoning behind the prediction, and clinicians are loath to use the thus-far predominately black-box technology. Both of the above points also have important implications for the certification of AI technology.

Therefore, in order to start harnessing the massive potential of machine learning for healthcare, and to actually use it to improve real patient outcomes, the Machine Learning in Medical Image Analysis group aims to do research that helps to bridge this gap between machine learning and clinical practice. We perform this research along four broad directions:

• Robustness, Safety and Uncertainty
• Interpretable Machine Learning
• Human-in-the-Loop Machine Learning Systems
• Generative Modelling on Big Medical Datasets

These topics are described in more detail in the research areas section below.

We are part of the Cluster of Excellence: Machine Learning - New Perspectives for Science and the University of Tübingen.

The group is headed by Dr. Christian Baumgartner.