From Unlabeled Data to Clinical Applications: Foundation Models in Medical Imaging

The performance of deep learning algorithms is highly dependent on the quantity and diversity of the available training data. However, obtaining sufficiently large datasets represents a significant challenge, particularly in the field of medical imaging. This study underscores the potential of self-supervised training strategies in the development of deep learning models for medical imaging tasks. It is demonstrated that workflows can be significantly optimized by incorporating the feature content of a large collection of medical X-ray images from intraoperative C-arm scans into a so-called foundation model. This approach facilitates the efficient adaptation to a variety of concrete applications by fine-tuning a small task-specific head network on top of the pre-trained foundation model, thereby reducing both computational demands and training time.