tags:
- medical
license: other
license_name: research-only-rail-m
model-index:
- name: Curia-MAE
results: []
extra_gated_prompt: >-
Please confirm that you have read and agree to the following disclaimer.
The model in this repository is provided for research use only (Research-only
RAIL-M license). The model(s) and/or software are not intended for use in
clinical decision-making or for any other clinical use, and performance for
clinical use has not been established.
🌐 Raidium | 📄 Paper Link | 🤗 Curia-2 | 🤗 Curia
Curia-MAE: Multi-Modal Multi-Anatomy MAE Pre-Training for 3D Medical Image Segmentation
We introduce Curia-MAE, a multi-modal, multi-anatomy Masked Autoencoder (MAE) pre-trained on roughly 300,000 CT and MRI volumes covering a large number of anatomical sites. Curia-MAE targets dense 3D medical image segmentation: it provides a strong, reusable frozen encoder that can be adapted to a new segmentation task by training only a lightweight decoder on top, while remaining competitive under full fine-tuning and particularly strong on lesion-focused tasks where labeled data is scarce. A single frozen encoder can thus be reused across diverse segmentation tasks, reducing the cost of adapting and deploying such models in clinical workflows.
Check the research paper: https://arxiv.org/abs/2608.05844
Finetuning with nnU-Net
Curia-MAE is a ResEnc-L (residual-encoder U-Net) backbone in the nnU-Net self-supervised pre-training format. Finetune it on a labeled 3D segmentation dataset with nnU-Net's finetuning-from-nnssl workflow:
# 1) dataset fingerprint + base plan
nnUNetv2_extract_fingerprint -d <DATASET_ID>
nnUNetv2_plan_experiment -d <DATASET_ID>
# 2) preprocess to the pre-training spacing/normalization and register this checkpoint
nnUNetv2_preprocess_like_nnssl -d <DATASET_ID> -n Curia-MAE \
-pc checkpoint_final.pth -am like_pretrained
# 3) finetune (the generated plan already embeds the checkpoint path)
nnUNetv2_train_pretrained <DATASET_ID> 3d_fullres <FOLD> \
-p ptPlans__Curia-MAE____Spacing__1.00_1.00_1.00___Norm__Z -tr PretrainedTrainer
Only the encoder and stem are transferred; the segmentation decoder is trained from scratch. The architecture and preprocessing plan are also provided in adaptation_plan.json.
License
The model is released under the RESEARCH-ONLY RAIL-M license. https://huggingface.co/raidium/curia/blob/main/LICENSE
Cite our paper
@article{danielou2026curiamae,
title={Curia-MAE: Multi-Modal Multi-Anatomy MAE Pre-Training for 3D Medical Image Segmentation},
author={Théo Danielou and Antoine Saporta and Léo Alberge and Corentin Dancette},
year={2026},
eprint={2608.05844},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2608.05844},
}
Acknowledgements
Curia-MAE uses the ResEnc-L residual-encoder U-Net architecture from nnU-Net v2 (Division of Medical Image Computing, German Cancer Research Center / DKFZ), and the released checkpoint is packaged to be fine-tunable within the nnU-Net framework. We gratefully acknowledge the nnU-Net project.