---
license: other
license_name: slimformer-plusplus-noncommercial-research-1.0
license_link: LICENSE
pipeline_tag: image-segmentation
tags:
  - pytorch
  - nnunet
  - medical-imaging
  - 3d-segmentation
  - slimformer-plusplus
  - mamba
  - arxiv:2610.25488
---

# SlimFormer++ — AMOS2022 CT

Official fold-0 pretrained checkpoint for **3D CT only medical image
segmentation**, released with
[SlimFormer++: Token Mixing Where It Matters for Efficient 3D Medical Image Segmentation](https://arxiv.org/abs/2610.25488).

[Source code](https://github.com/WHHHHHY822/SlimFormerPlusPlus) · [Paper](https://arxiv.org/abs/2610.25488) ·
[Hugging Face Papers](https://huggingface.co/papers/2610.25488) ·
[Original checkpoint release](https://github.com/WHHHHHY822/SlimFormerPlusPlus/releases/tag/amos2022-ct-best-v1) ·
[Dataset](https://amos22.grand-challenge.org/)

This repository mirrors the original inference-only checkpoint **without
changing the selected network weights**, along with its nnU-Net plans and
dataset metadata. Training images, segmentation masks, and optimizer state
are not distributed here.

## Model

SlimFormer++ is a custom nnU-Net v2 trainer. It uses windowed Mamba mixing in
the first encoder stage, residual channel MLPs in deeper stages, and a
depthwise-convolution bottleneck augmented with sparse long-range context.
This checkpoint is for the original Mamba-based model. The separate Mamba-free
implementation has no released pretrained checkpoint.

| Setting | Value |
| --- | --- |
| Dataset | AMOS2022 CT (`Dataset004_AMOS`) |
| Input | One CT only channel, NIfTI `.nii.gz` |
| Trainer | `nnUNetTrainer_SlimFormerPlusPlus` |
| Configuration / fold | `3d_fullres` / `0` |
| Selected checkpoint epoch | 1199 |
| Completed training epochs | 1200 |
| Training / validation cases | 200 / 100 |
| Inference patch | `[64, 192, 160]` |

### Output labels

| Label ID | Structure |
| --- | --- |
| 0 | background |
| 1 | spleen |
| 2 | right kidney |
| 3 | left kidney |
| 4 | gallbladder |
| 5 | esophagus |
| 6 | liver |
| 7 | stomach |
| 8 | aorta |
| 9 | inferior vena cava |
| 10 | pancreas |
| 11 | right adrenal gland |
| 12 | left adrenal gland |
| 13 | duodenum |
| 14 | bladder |
| 15 | prostate/uterus |

## Evaluation

The paper's Table III results and the released checkpoint's fold-0 validation
measurements are reported separately. They are different evaluation records.
The fold-0 results are **not independent test-set results**.

| Evaluation record | Mean Dice (%) | Mean HD95 (mm) |
| --- | ---: | ---: |
| Paper, Table III | 89.76 | 11.36 |
| Released checkpoint, fold-0 validation (100 cases) | 90.4005 | 10.0117 |

The validation aggregates are copied from the project's published
[reproduction materials](https://github.com/WHHHHHY822/SlimFormerPlusPlus/blob/92e66fcd8428bd9aeae94e44df36842078c60006/experiments/amos-ct/metrics.json).
The [protocol](reproduction/protocol.json) records training and inference
settings; [metrics](reproduction/metrics.json) contains the reference
aggregates. See the [reproduction guide](https://github.com/WHHHHHY822/SlimFormerPlusPlus/blob/92e66fcd8428bd9aeae94e44df36842078c60006/experiments/README.md)
for data preparation, original splits, per-case results and evaluation commands.

## Use the checkpoint

1. Follow the project's [installation guide](https://github.com/WHHHHHY822/SlimFormerPlusPlus/blob/92e66fcd8428bd9aeae94e44df36842078c60006/docs/INSTALLATION.md).
   The reference setup uses Linux, an NVIDIA GPU, Python 3.12, PyTorch 2.6.0
   and CUDA Toolkit 12.6. Mamba and causal-conv1d are compiled from source.
   Source revision used for this model card: `92e66fcd8428bd9aeae94e44df36842078c60006`.
2. Install `huggingface_hub` in that environment and download this repository
   into a fresh nnU-Net results directory:

```bash
python -m pip install huggingface_hub
export nnUNet_results=/path/to/empty/nnUNet_results
hf download Yang0822/SlimFormerPlusPlus-AMOS2022-CT --local-dir "$nnUNet_results"
(cd "$nnUNet_results" && sha256sum -c SHA256SUMS)
```

3. Prepare one input volume per case as `CASE_ID_0000.nii.gz`, using the
   matching CT only modality, and run:

```bash
export nnUNet_compile=false
nnUNetv2_predict -i /path/to/imagesTs -o /path/to/predictions \
    -d 4 -c 3d_fullres \
    -tr nnUNetTrainer_SlimFormerPlusPlus -f 0 \
    -chk checkpoint_best.pth -step_size 0.5
```

The original dataset/model directory structure is preserved so the nnU-Net
predictor can discover the checkpoint. Downloading files does not install the
trainer or its CUDA dependencies. This model uses the project's custom nnU-Net
runtime; see the commands above for inference.

## Intended use and limitations

The release is intended for noncommercial scientific research and reproducibility.
Its validation results describe the published dataset split and do not establish
performance on other scanners, institutions, populations or imaging protocols.
This release does not establish suitability for clinical diagnosis or deployment.
Dataset access and use remain subject to the dataset providers' terms.

## License and provenance

**Noncommercial research use only.** The existing
[SlimFormer++ Noncommercial Research License 1.0](LICENSE) is preserved.
Third-party components retain their own licenses and notices; see
[NOTICE.md](NOTICE.md) and [Apache-2.0](licenses/Apache-2.0.txt).

[provenance.json](provenance.json) records the original release URL, archive
SHA256, checkpoint SHA256 and source revision. `SHA256SUMS` is the unmodified
checksum list from the original GitHub checkpoint archive.

## Citation

Please cite SlimFormer++ together with nnU-Net:

```bibtex
@misc{hong2026slimformertokenmixingmatters,
      title={SlimFormer++: Token Mixing Where It Matters for Efficient 3D Medical Image Segmentation},
      author={Yang Hong and Lei Zhang and Zhihong Liu and Jianqing Mo and Xujiong Ye},
      year={2026},
      eprint={2610.25488},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2610.25488},
}
```

[nnU-Net source and citation](https://github.com/MIC-DKFZ/nnUNet)