Making AI ready for clinical practice

Prof. Lars French's opening speech was followed by PD Dr. Tobias Lasser presentation on the latest findings and developments of his TUM research group “Computational Imaging and Inverse Problems” in the context of clinical decision support in radiology and dermatology respectively. While AI algorithms already excel in correctly identifying diseases in certain application scenarios, such as lung diseases in X-ray radiographs or skin diseases in photographs, several elements for daily clinical usage are still missing representing the focus of Dr. Lasser’s research group. One example are novel fusion schemes merging images with additional patient information for better diagnostic quality. Another example was the proof that tiny image details that are routinely discarded have – not surprisingly - high relevance on algorithms’ outcomes.

Yet another example is related to the interpretability of the algorithms’ outcomes, so that doctors better understand how to apply algorithms in clinical practice, thereby building trust and providing support. Last but not least patient data can be excluded for which algorithms cannot make an accurate prediction due to limitations in training data set. Alessandro Wollek, a doctoral candidate in Dr. Lasser’s research group, presented several of those novel findings in detail during his presentation.

Dr. Sebastian Krammer’s presentation outlined the group's research studies on detection of skin diseases using image analysis. Stressing how digitization with so-called software Avelios contributed to documentation efficiency at LMU University Hospital, Dr. Krammer exemplified up to 2,000 structured data points in case of a single patient treatment. Based on this meta data abundance, a new algorithm for detecting skin diseases has been developed characterized by a multimodal, multistep data integration process.

Prof. Michael Ingrisch, Director of the LMU Clinical Data Science Group, is convinced that asking the right questions - clinical questions with clinical relevance - is crucial. Above all, it would be important that these questions are answered in a clinical context, e.g. by using AI systems on critical subgroups.

PD. Dr. Bastian Sabel discussed gradual implementation of AI in radiology, i.e. phase by phase, in his lecture. Despite the advantages of this technology, he emphasized the need to address ethical and practical concerns, especially regarding data privacy and the role of human expertise in an increasingly digital medical landscape.

Prof. Alena Buyx's TUM Embedded Ethics team reported on how they embedded ethics within the consortium’s project setting. As of now, “ethics” is often perceived as an afterthought. Therefore, the team’s mission is to bring ethics “to the workbench”, for example by discussing ethical issues as soon as they arise. In the context of the DR-AI project, the aforementioned approach was implemented by conducting an interview study with all participants and carrying out ethnographic fieldwork. The study’s results stress researcher’s responsibility in machine learning development processes and, contribute to high-quality decision-making in medicine. Besides offering insights into biomedical ethics research, the team’s work package provided an early example of successful embedded ethics application and yielded essentials tailoring the approach closer to computer scientists’and clinicians’needs.