GRK 3012:  KEMAI - Knowledge Infusion and Extraction for Explainable Medical AI

• Subject Area: Computer Science ; Medicine
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 520750254

Since decades, computer aided methods support radiologists and other medical experts with functionalities such as segmentation, registration, or quantification. In contrast to such systems that contribute a set of basic functionalities to the diagnosis process, expert systems aim for a broader functionality and can perform the entire diagnosis, or even treatment planning. While the employed knowledge base helps to provide explanations for these expert systems' recommendations, they require extensive work to curate the knowledge base. Modern deep learning systems promise to address this downside, since they do not require a curated knowledge base. At the same time such systems enable new accuracy heights, which have not been imagined before. Learning systems come, however, with their own set of challenges. It is not only, that the decision formed by the learning system might deviate from the decision of a medical expert, sometimes considerably, but also that the reasons for this decision can usually not be comprehended, as learning systems are not inherently explainable. As a consequence, it is difficult for medical doctors who are confronted with these predictions, to probe and trust them, to defend them against other experts, or even justify them to patients. Unfortunately, current research and the thereof derived medical learning systems often rather focus on prediction accuracy, as measured against a chosen test set, than on the understanding and communication of the obtained predictions. Therefore, within this research training group, we aim at combining the benefits of knowledge- and learning-based systems, to not only allow for state of the art accuracy in medical diagnosis, but to also clearly communicate the obtained predictions to medical doctors, while also addressing the ethical implications within the medical decision process. Furthermore, by infusing the knowledge contained in medical guidelines and other documents into a machine learning model, we plan to also tackle the challenges which come with data scarcity and many features. To work towards our goals, we have identified four PhD topic areas: data augmentation, knowledge infusion into and knowledge extraction out of learning systems, as well as model explanation. In each of these areas several potential PhD topics have been identified. In order to do tackle the addressed challenges, we closely interweave the competences from computer science, medicine, and ethics, brought in by the participating researchers, which form interdisciplinary thesis advisory committees for each PhD topic.

• DFG Programme: Research Training Groups
• Applicant Institution: Universität Ulm
• Spokesperson: Professor Dr. Timo Ropinski
• Participating Researchers: Professor Dr. Ambros Beer; Professor Dr. Meinrad Beer; Professor Dr. Daniel Alexander Braun; Professorin Dr. Birte Glimm; Professor Dr.-Ing. Michael Götz; Professorin Dr. Rebekka Hufendiek; Professor Dr. Hans A. Kestler; Dr. Catharina Lisson; Professor Dr. Heiko Neumann; Professor Dr. Ansgar Scherp; Professor Dr. Florian Steger; Dr. Ina Vernikouskaya