The goal of this project is to reach a new frontier in precision radiology by developing foundation models for 3D medical imaging, trained with self-supervised learning (SSL) techniques at an unprecedented scale. Building on methods like DINOv2 and CLIP, this work will leverage a uniquely large and diverse dataset comprising over 150,000 anonymised CT and MRI exams, totaling more than 130 TB of data.
This project addresses one of the key limitations in medical AI - the dependence on large-scale annotated datasets, which are difficult and expensive to obtain in clinical practice. By scaling SSL approaches to 3D data, this project aims to learn rich representations of anatomy and pathology without requiring manual labels which is a critical step toward enabling general-purpose medical vision models.
The scientific contribution of this work lies in the adaptation and scaling of SSL methods specifically for volumetric, multimodal data. It will demonstrate the ability of 3D foundation models to transfer across a wide range of downstream tasks, such as disease classification, lesion detection, or prognosis, with limited supervision. This approach has the potential to significantly improve diagnostic accuracy, reduce radiologists’ workload, and advance personalised medicine through more consistent and data-driven interpretation.
This initiative pushes the boundaries of what’s currently possible in medical computer vision and AI for healthcare. A research outcome will include publishing project results in a high-impact journal such as Nature Medicine, showcasing the feasibility and value of large-scale 3D SSL. The project researchers also plan to release some model weights, supporting transparency, reproducibility, and further research across the medical AI community.
Antoine Saporta, Raidium, France