About this role
Summary: Join the R&D team to develop generative vision-language models for radiology, focusing on foundation model pre-training and adaptation. You will design experiments, train models on large-scale medical datasets, and collaborate with radiologists to improve diagnostic tools.
Foundation models are pre-trained on large amounts of data and can be adapted to many downstream tasks with few or no labels. At Raidium, we build our own foundation models for radiology, for 3D CT and MRI imaging.
We released [Curia](https://arxiv.org/abs/2509.06830), a state-of-the-art vision encoder for CT and MRI, trained on about 150,000 clinical exams. We then built [Jolia](https://arxiv.org/abs/2606.24570), a vision-language model for 3D CT that learns from radiology reports at the level of individual findings. Our next steps include localization, multi-volume and longitudinal reasoning, and more interactive use of these models by radiologists.
As an intern, you join the research team and work on one research topic related to foundation model pre-training and adaptation. In particular, you will work development of generative VLMs with image grounding and localisation, with applications to findings classification and evolution of pathologies through time.
In this internship, you will:
- Read the latest literature on foundation models research for vision and language
- Design and run experiments on 3D vision and vision-language foundation models for radiology
- Train models at scale on Raidium's dataset of radiology exams and reports, using GPU clusters
- Benchmark your approach on clinical downstream tasks (detection, segmentation, classification, report generation)
- Work with radiologists to define the needs and to check the results
- Share your results with the team, and when possible, publish them in a top-tier conference or journal
You will have a daily follow-up with a researcher from the team, and a weekly meeting with senior researchers. The internship will be managed by a lead researcher at Raidium.
### About Raidium
Raidium builds foundation models for radiology and the software that brings them to radiologists. Our models, Curia and Jolia, power a new generation of tools to read and analyze CT and MRI exams.
Raidium was founded by experts in radiology and machine learning:
- Pierre Manceron, CEO, Centrale Paris engineer in Applied Mathematics and Machine Learning, former ML lead at Owkin
- Paul Herent, CMO, radiologist and former interne des hôpitaux de Paris, Cogmaster (ENS) and PSL, with published research on deep learning in radiology
### Your profile
- Final year of MSc in Machine Learning, Computer Science, Mathematics or Electrical Engineering (for example MVA), or PhD student looking for a research internship
- Strong knowledge of Deep Learning, with Computer Vision and/or NLP
- Good Python skills, and experience with PyTorch
- Interest in research and in the healthcare domain
- Authorization to work in France
### Pluses (not required):
- Experience with Transformers, self-supervised learning or vision-language models
- Experience with 3D or medical images
- Experience with version control such as git
- Experience with HPC clusters and multi-GPU training
- Experience with coding agents such as Claude Code or Codex
- Publications in top-tier Machine Learning or medical imaging venues
### Why join us
- Work on real research problems with a direct impact on patient care
- Train large models on a unique dataset of radiology exams and reports (130 TB)
- Access to large GPU clusters, including European supercomputers
- Work with a team of expert ML researchers, engineers and radiologists
- Apply and improve the latest Deep Learning technologies to build an inspiring product
- Have a real and positive impact on people's lives
- Access to premium computing hardware
- Competitive salary