datasets:
- AnonRes/OpenMind
license: cc-by-4.0
pipeline_tag: image-segmentation
tags:
- medical
Primus: Enforcing Attention Usage for 3D Medical Image Segmentation
This repository contains pre-trained checkpoints for Primus, a Transformer-centric architecture for 3D medical image segmentation.
- Paper: Primus: Enforcing Attention Usage for 3D Medical Image Segmentation
- Code: MIC-DKFZ/nnUNet
- Pre-training codebase: MIC-DKFZ/nnssl
OpenMind Benchmark 3D SSL Models
Model from the paper: An OpenMind for 3D medical vision self-supervised learning
Pre-training codebase used to create checkpoint: MIC-DKFZ/nnssl
Dataset: AnonRes/OpenMind
Downstream (segmentation) fine-tuning: TaWald/nnUNet
Overview
This repository hosts pre-trained checkpoints from the OpenMind benchmark:
📄 An OpenMind for 3D medical vision self-supervised learning (Wald, T., Ulrich, C., Suprijadi, J., Ziegler, S., Nohel, M., Peretzke, R., ... & Maier-Hein, K. H. (2024).)
(arXiv:2412.17041) — the first extensive benchmark study for self-supervised learning (SSL) on 3D medical imaging data.
Each model was pre-trained using a particular SSL method on the OpenMind Dataset, a large-scale, standardized collection of public brain MRI datasets.
These models are not recommended to be used as-is for feature extraction. Instead we recommend using the downstream fine-tuning frameworks for segmentation and classification adaptation, available in the adaptation repository. While manual download is possible, we recommend using the auto-download feature of the fine-tuning repository by providing the repository URL on Hugging Face instead of a local checkpoint path.
Model Variants
We release SSL checkpoints for two backbone architectures:
- ResEnc-L: A CNN-based encoder [a, b]
- Primus-M: A transformer-based encoder introduced in the Primus paper
Each encoder has been pre-trained using one of the following SSL techniques:
| Method | Description |
|---|---|
| Volume Contrastive (VoCo) | Contrastive pretraining method for 3D volumes |
| VolumeFusion (VF) | Spatial volume fusion-based segmentation SSL method |
| Models Genesis (MG) | Reconstruction and denoising based pretraining method |
| Masked Autoencoders (MAE) | Default reconstruction based pretraining method |
| Spark 3D (S3D) | Sparse reconstruction based pretraining mehtod (CNN only) |
| SimMIM | Simple masked reconstruction based pretraining method (TR only) |
| SwinUNETR SSL | Rotation, Contrastive and Reconstruction based pre-training method. |
| SimCLR | Transfer of 2D Contrastive learning baseline method to 3D |
Citation
If you find this model or the OpenMind benchmark useful, please cite:
@article{wald2025primus,
title={Primus: Enforcing Attention Usage for 3D Medical Image Segmentation},
author={Wald, Tassilo and Roy, Saikat and Isensee, Fabian and Ulrich, Constantin and Ziegler, Sebastian and Trofimova, Dasha and Stock, Raphael and Baumgartner, Michael and K{\"o}hler, Gregor and Maier-Hein, Klaus},
journal={arXiv preprint arXiv:2503.01835},
year={2025}
}
@article{wald2024openmind,
title={An OpenMind for 3D medical vision self-supervised learning},
author={Wald, Tassilo and Ulrich, Constantin and Suprijadi, J. and Ziegler, Sebastian and Nohel, M. and Peretzke, R. and others},
journal={arXiv preprint arXiv:2412.17041},
year={2024}
}