Julius Hense

PhD project:  Responsible Machine Learning for Multimodal Medical Data

Machine learning has the potential to revolutionize healthcare. However, its clinical adoption is still hampered by various roadblocks, including insufficient model robustness, data efficiency, and explainability. Furthermore, medical domains often pose specific challenges that established machine learning algorithms do not account for. My research is focused on contributing to overcoming these challenges, aiming at making machine learning in healthcare more responsible. I am particularly interested in learning from multimodal medical data. More specifically, I aim to build ML systems that combine the analysis of medical images with other patient-centred data, such as electronic health records, time-series, or multi-omics data, and utilize them for downstream tasks like screening, diagnosis, or biomarker discovery. For that purpose, I work with techniques from multimodal machine learning, representation learning, and explainable AI, e.g., to design targeted multimodal fusion methods, compute low-dimensional representations of various medical data modalities, or uncover biological patterns across modalities. I specialize in applications for digital pathology and oncology, where I have the chance to collaborate with leading domain experts.