---

🌐 Blog Post | 📄 Curia-2 Paper Link | 🤗 Original Curia | 📄 Original 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.

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}, 
}