---

🌐 Blog Post | 🤗 Original Curia | 📄 Curia Paper Link

Curia-2: Scaling Self-Supervised Learning for Radiology Foundation Models

We introduce Curia-2, a follow-up to Curia which significantly improves the original pre-training strategy and representation quality to better capture the specificities of radiological data. Curia-2 excels on vision-focused tasks and fairs competitively to vision-language models on clinically complex tasks such as finding detection.

Research paper coming soon.

Loading the model

To load the model, use the AutoModel class from huggingface transformers library.

from transformers import AutoModel
model = AutoModel.from_pretrained("raidium/curia-2")

You can also load the image pre-processor

from transformers import AutoImageProcessor
processor = AutoImageProcessor.from_pretrained("raidium/curia-2", trust_remote_code=True)

Then to forward an image:

img = 2048 * np.random.rand(256, 256) - 1024 # single axial slice, in PL orientation
model_input = processor(img)
features = model(**model_input)

The image must follow the following format:

input: numpy array of shape (H, W)
  Images needs to be in:
  - PL for axial
  - IL for coronal
  - IP for sagittal
  for CT, no windowing, just hounsfield or normalized image
  for MRI, similar, no windowing, just raw values or normalized image

License

The model is released under the RESEARCH-ONLY RAIL-M license. https://huggingface.co/raidium/curia/blob/main/LICENSE

Cite our paper

@article{saporta2026curia2,
      title={Curia-2: Scaling Self-Supervised Learning for Radiology Foundation Models}, 
      author={Antoine Saporta and Baptiste Callard and Corentin Dancette and Julien Khlaut and Charles Corbière and Leo Butsanets and Amaury Prat and Pierre Manceron},
      year={2026},
      eprint={2604.01987},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2604.01987}, 
}