---
datasets:
- AnonRes/OpenMind
license: cc-by-4.0
pipeline_tag: image-segmentation
tags:
- medical
---

# OpenMind Benchmark 3D SSL Models: Primus-M

This repository hosts pre-trained checkpoints for the **Primus-M** backbone architecture, as introduced in the paper:

📄 **Primus: Enforcing Attention Usage for 3D Medical Image Segmentation**  
(Wald, T., Roy, S., Isensee, F., Ulrich, C., Ziegler, S., Trofimova, D., Stock, R., Baumgartner, M., Köhler, G., & Maier-Hein, K.)  
[[Hugging Face Paper](https://huggingface.co/papers/2503.01835)] [[arXiv:2503.01835](https://arxiv.org/abs/2503.01835)]

These weights were also utilized in the **OpenMind** benchmark study:  
📄 **An OpenMind for 3D medical vision self-supervised learning**  
[[arXiv:2412.17041](https://arxiv.org/abs/2412.17041)]

---

![OpenMind](https://huggingface.co/datasets/AnonRes/OpenMind/resolve/main/assets/OpenMindDataset.png)

## Overview

This repository hosts pre-trained checkpoints from the **OpenMind** benchmark, 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](https://huggingface.co/datasets/AnonRes/OpenMind), a large-scale, standardized collection of public brain MRI datasets. 

This specific checkpoint uses the **Primus-M** transformer-centric encoder pre-trained with the **VolumeFusion (VF)** self-supervised method. Primus addresses the over-reliance on convolutional blocks in existing architectures, leveraging high-resolution tokens and advanced block design to achieve state-of-the-art performance in 3D medical image segmentation.

**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](https://github.com/TaWald/nnUNet) and the main [nnU-Net](https://github.com/MIC-DKFZ/nnUNet) framework.

## Model Variants

We release SSL checkpoints for two backbone architectures:

- **ResEnc-L**: A CNN-based encoder [[a](https://arxiv.org/abs/2410.23132), [b](https://arxiv.org/abs/2404.09556)]  
- **Primus-M**: A transformer-based encoder [[Primus paper](https://arxiv.org/abs/2503.01835)]

Each encoder has been pre-trained using one of the following SSL techniques: Volume Contrastive (VoCo), VolumeFusion (VF), Models Genesis (MG), Masked Autoencoders (MAE), Spark 3D (S3D), SimMIM, SwinUNETR SSL, or SimCLR.

## Implementation

The code for Primus and the fine-tuning pipeline is available in the official nnU-Net repository:
- **Code**: [MIC-DKFZ/nnUNet](https://github.com/MIC-DKFZ/nnUNet/blob/master/documentation/primus.md)
- **Pre-training codebase**: [MIC-DKFZ/nnssl](https://github.com/MIC-DKFZ/nnssl)

## Citation

If you use this model, please cite the following:

```bibtex
@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 Köhler, 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}
}
```