SummaryArtificial intelligence (AI) methods have repeatedly shown powerful modeling capabilities and pattern recognition across a wide range of tasks and domains approximating and sometimes even surpassing that of human experts, with one such domain being medical imaging. However, the tradeoff for such powerful models is the requirement of large-scale, annotated datasets to leverage for training and tuning the models. Due to patient recruitment being limited by study constraints, most studies cannot accrue a sufficient body of patients to produce robust and generalisable models that can be properly validated, hampering the translatability of AI into clinical routine. One method of tackling such a hurdle towards clinical AI implementation relies on the use of self-supervised learning (SSL) to create models from large-scale multicentric datasets that do not per se require annotated data, thus producing a flexible feature extractor, or foundation model, that can then be deployed on smaller datasets for modeling rather than training a network from scratch. Due to the lack of publicly available SSL models in the radiology domain and the national scale of the study, the aim of the project is to leverage the NAKO cohort MRI data in order to first train SSL models that do not require labels, and then to use such models to train task-specific models such as for correlates of psychiatric diseases or stroke with brain MRI or for cardiovascular endpoints such as risk of myocardial infarction using heart MRI.

Keywords Artificial-Intelligence Multimodal Self-supervised-models

InstitutionsTU Dresden, Radiologie Freiburg, Else Kröner Fresenius Center for Digital Health, Technische Universität Dresden, Technische Universität Dresden, EKFZ- Else Kröner Fresenius Zentrum Dresden